A cognitive engagement tracking method based on dual features and semi-supervised learning

Abstract

The present invention relates to the fields of natural language processing and educational data mining, and provides a cognitive investment tracking method based on dual features and semi-supervised learning, including collecting and cleaning the discussion text data of learners in an online learning platform, and marking 10%-40% of the discussion text data; performing text enhancement operations on the unlabeled discussion data to obtain enhanced unlabeled text data; extracting the BERT semantic and LIWC cognitive dual features of all discussion text data; fusing the dual features of each discussion text to obtain a fused new feature; and training a cognitive investment tracking model using text data with dual features. The method of the present invention captures discussion text features from both cognitive and semantic aspects, and uses a small amount of labeled text data and a large amount of unlabeled text data to train a cognitive investment tracking model, effectively helping teachers grasp the level of cognitive investment of learners.

Description

Technical Field

[0001] The present invention relates to the fields of natural language processing and educational data mining, and in particular to a cognitive engagement tracking method based on dual features and semi-supervised learning. Background Art

[0002] Cognitive engagement refers to the choice of learning methods and strategies adopted by learners to ensure the achievement of learning goals, as well as the monitoring and regulation of the entire learning activity. Cognitive engagement in learners' interactive discourse is positively correlated with their learning outcomes. Cognitive engagement tracking is a study that assesses learners' cognitive engagement levels in interactive discussions. Cognitive engagement tracking for interactive discourse data provides new possibilities for in-depth insights into the cognitive development trends of individuals and groups and for optimizing the learning process.

[0003] There are two main areas of cognitive engagement tracking:

[0004] (1) Cognitive engagement tracking method based on supervised learning

[0005] The cognitive engagement tracking method based on supervised learning is a machine learning task that learns a cognitive engagement tracking model from a large amount of labeled text training data to infer the level of cognitive engagement implicit in the discussion text. The cognitive engagement tracking model generated based on the supervised learning training method consists of a cognitive feature representation module and a cognitive reasoning module, which relies on a large number of samples of known categories to adjust the parameters of the cognitive engagement tracking model.

[0006] (2) Cognitive engagement tracking method based on semi-supervised learning

[0007] The cognitive engagement tracking method based on semi-supervised learning uses a large amount of unlabeled discussion text data and a small amount of labeled discussion text data, and trains the cognitive engagement tracking model through label propagation between labeled data and unlabeled data to complete the cognitive engagement tracking task. Most of the currently commonly used cognitive engagement tracking methods based on semi-supervised learning infer different levels of cognitive engagement from the semantic level of discussion text.

[0008] Although the above-mentioned cognitive engagement tracking methods have shown good results in past applications, they all have their own shortcomings. The cognitive engagement tracking methods based on supervised learning have low reasoning accuracy when the amount of labeled data is small. Training a cognitive engagement tracking model with superior performance often requires relying on a large amount of labeled data. However, labeled data is often difficult to obtain, and the data labeling process is time-consuming and laborious. The cognitive engagement tracking methods based on semi-supervised learning can use a small amount of labeled text data and a large amount of unlabeled text data for reasoning, but most cognitive engagement tracking models based on semi-supervised learning only reason from the semantic level of text data, ignoring the impact of the cognitive level characteristics of text data on the reasoning results. Summary of the invention

[0009] In view of the above defects or improvement needs of the prior art, the present invention provides a cognitive engagement tracking method based on dual features and semi-supervised learning. The method captures discussion text features from both cognitive and semantic aspects, and uses a small amount of annotated text data and a large amount of unannotated text data to train a cognitive engagement tracking model. The method can more accurately obtain the cognitive engagement state contained in the learner's interactive discourse, thereby helping teachers to intervene in learners' behaviors more targeted and provide support for optimizing the learning process.

[0010] The objectives of the present invention are achieved through the following technical solutions.

[0011] A cognitive engagement tracking method based on dual features and semi-supervised learning includes the following steps:

[0012] (1) Collect and clean the discussion text data of learners in the online learning platform, and mark a small part of the discussion text data as labeled text data;

[0013] (2) Perform text enhancement operations on the unlabeled discussion data to obtain enhanced unlabeled text data, completing the construction of the labeled dataset, unlabeled dataset, and enhanced unlabeled dataset;

[0014] (3) Extract BERT semantic and LIWC (Linguistic Inquiry and Word Count) cognitive dual features from all discussion text data in the labeled dataset, unlabeled dataset, and enhanced unlabeled dataset;

[0015] (4) Fuse the dual features of each discussion text to obtain new fused features;

[0016] (5) Use labeled datasets with dual features, unlabeled datasets, and enhanced unlabeled datasets to train the cognitive engagement tracking model.

[0017] It can be seen from the technical solution provided by the present invention that the method does not require a large amount of annotated text data, and only a small amount of annotated text and a large amount of unannotated text can be used to train the cognitive engagement tracking model, which reduces the time and manpower consumption of annotating text. At the same time, the method infers the cognitive engagement level of text data from the two dimensions of semantics and cognition, making up for the disadvantage of the unidimensionality of the traditional cognitive engagement tracking model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A flowchart of the cognitive engagement tracking method based on dual features and semi-supervised learning provided by the present invention.

[0019] Figure 2 This is a diagram of the training architecture of the cognitive engagement tracking model provided by the present invention. DETAILED DESCRIPTION

[0020] In order to more clearly and completely describe the purpose and technical solution of the present invention, the details of the present invention are described in detail below in conjunction with the drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the protection scope of the present invention.

[0021] See also Figure 1 , Figure 2 As shown, an embodiment of the present invention provides a cognitive engagement tracking method based on dual features and semi-supervised learning, comprising the following steps:

[0022] A000: Collect and clean the discussion text data of learners in the online learning platform, and mark 10%-40% of the discussion text data as labeled text data.

[0023] In this embodiment, 7510 discussion texts of learners in the cloud classroom course "Foundations of Psychology" of Central China Normal University are first obtained by crawling.

[0024] Then, the data cleaning operation is completed based on Python. This operation includes using the re module in Python to edit regular expressions to match and delete dirty data such as HTML tags and irrelevant characters in the initial discussion text data, using the hanziconv module to simplify the traditional Chinese text in the text data, and matching and removing stop words in the text data based on the stop word dictionary of Harbin Institute of Technology.

[0025] Finally, 10%-40% of all the data are annotated manually. In this embodiment, in order to test the performance of the cognitive engagement tracking model, all text data are annotated and divided into a training set and a test set at a ratio of 9:1. Among them, 10% of the data in the training set is selected as the annotated data, and the remaining 90% of the data is ignored and used as unlabeled data.

[0026] A001: Perform text enhancement operations on unlabeled discussion data to obtain enhanced unlabeled text data, and complete the construction of labeled datasets, unlabeled datasets, and enhanced unlabeled datasets.

[0027] In this embodiment, first, English is used as the intermediate language to perform a reverse translation of "Chinese-English-Chinese" on each unlabeled discussion text to obtain text data with the same semantics as the unlabeled text.

[0028] Then, each text data after reverse translation is first segmented to obtain the text representation X = {x 1 ,x 2 ,…,x N}, where x i represents the word at the i-th position in the text, and N represents the number of words in the text. Then calculate the probability of each word in the text being replaced Where p is a hyperparameter, TFIDF(x i ) is the TFIDF value of the i-th word, max i TFIDF(x i ) is the maximum TFIDF value in sentence X. In each piece of text data, each word in the text is sampled according to its corresponding replacement probability T to determine whether the word is replaced with a TF-IDF non-core word.

[0029] Finally, all the reverse-translated texts are segmented and a replacement dictionary is constructed based on the segmentation results. The frequency of each word in the dictionary is For words in the sentence that need to be replaced by TF-IDF non-core words, a word is extracted from the dictionary according to the probability p(w) for replacement, and finally the enhanced unlabeled text data is obtained. Among them, each unlabeled text corresponds to an enhanced unlabeled text data.

[0030] A002: Extract BERT semantic and LIWC cognitive dual features of all discussion text data in labeled datasets, unlabeled datasets, and enhanced unlabeled datasets.

[0031] In this embodiment, the feature representation of each text data is constructed from two levels: BERT semantic features and LIWC cognitive features. Semantic features and cognitive features complement each other in text feature representation, providing a richer feature representation for the cognitive input tracking model.

[0032] When extracting BERT semantic features from text data, the bert-base-chinese version model is used, which is a semantic feature representation model pre-trained based on general corpora such as Chinese Wikipedia. The model has 12 hidden layers, and the dimension of each word vector in the hidden layer is 768 dimensions.

[0033] The specific steps of extracting BERT semantic features are as follows: each discussion text in all data sets is input into the BERT model based on Chinese prediction pre-training, and the first vector in the last layer output sequence is selected as the semantic feature representation of the discussion text, that is: E bert ={e 1 ,e 2 ,…,e n}. Where n is the dimension of the hidden layer in the BERT model.

[0034] When extracting LIWC cognitive features of text data, the TextMind Chinese psychological analysis system from the Computational Network Psychology Laboratory of the Institute of Psychology, Chinese Academy of Sciences, was used. The TextMind system extracts 102-dimensional features included in text data by integrating the Chinese LIWC dictionary. The system captures the statistical frequency features of different categories of words through a large number of regular expressions based on the Chinese LIWC dictionary, providing a basis for distinguishing different levels of cognitive investment.

[0035] The specific steps of extracting LIWC are as follows: each discussion text in all data sets is input into the LIWC cognitive dictionary, and the LIWC cognitive feature E of each discussion text is calculated according to the following formula: LIWC ={e 1 ,e 2 ,…,e m}.

[0036]

[0037] Among them, s is the sth LIWC cognitive feature, Dic s The corresponding dictionary.

[0038] A003: Fuse the dual features of each discussion text to obtain the fused new features.

[0039] The specific steps of integrating BERT semantic features and LIWC cognitive features are as follows: bert and LIWC cognitive characteristics E LIWC Linear splicing to generate an intermediate fusion feature E of length m+n BL ={e 1 ,…,e n ,…,e m+n}, then the intermediate fusion feature is input into the linear connection layer according to the following formula, and then passes through the Tanh activation function to output the final fusion feature E con ={e 1 ,e 2 ,…,e k}.

[0040] E con =Tanh(W 1 ·E BL +b 1 )

[0041] Among them, W 1 is the weight matrix of the linear connection layer, b 1 represents the bias of the linear connection layer, and k is the output feature E after feature fusion. con Dimension.

[0042] A004: Train a cognitive engagement tracking model using labeled datasets with dual features, unlabeled datasets, and enhanced unlabeled datasets.

[0043] First, the fusion features of all discussion texts are put into the fully connected layer, and the softmax activation function is applied to the output of the fully connected layer to obtain the cognitive investment distribution corresponding to each discussion text. Then different calculations are performed based on different data sets.

[0044] For a labeled data set, the loss function between the labeled data and the corresponding label is calculated according to the following formula:

[0045]

[0046] Among them, y is the label of the labeled data, p l is the cognitive input distribution of the labeled data. For the unlabeled data set and the enhanced unlabeled data set, the cognitive input distribution of the enhanced unlabeled data needs to be sharpened first. Then the loss function of the unlabeled data and the corresponding enhanced unlabeled data is calculated according to the following formula:

[0047]

[0048] Among them, p u is the cognitive input distribution of unlabeled data, It is the sharpened cognitive input distribution of the enhanced unlabeled data.

[0049] Finally, the two loss functions are used to calculate the objective function according to the following formula:

[0050] Loss Final =Loss Supervised +λLoss Consistency

[0051] Among them, λ is a hyperparameter used to control the impact of unlabeled data on model parameters. The parameters of the cognitive engagement tracking model are updated by minimizing the objective function to complete the training of the cognitive engagement tracking model.

[0052] The above scheme in this embodiment aims to solve the problems of low reasoning accuracy of previous cognitive engagement tracking methods in the absence of labeled data and lack of cognitive-level feature representation of text data by constructing a cognitive engagement tracking model using a cognitive engagement tracking method based on semi-supervised learning and BERT semantic and LIWC cognitive dual features, thereby solving the above problems.

[0053] The contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0054] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A cognitive engagement tracking method based on dual features and semi-supervised learning, characterized in that: The following steps are involved: (1) Collect and clean the discussion text data of learners in the online learning platform, and mark 10%-40% of the discussion text data as labeled text data; (2) Perform text enhancement operations on the unlabeled discussion data to obtain enhanced unlabeled text data, and complete the construction of the labeled data set, the unlabeled data set, and the enhanced unlabeled data set; specifically, the following steps are included: (2-1) Using English as the intermediate language, perform a reverse translation of "Chinese-English-Chinese" on each unlabeled discussion text to obtain text data with the same semantics as the unlabeled text; (2-2) Each text data after reverse translation is first segmented to obtain the segmented text representation X = {x1, x2, ..., x N }, where x i represents the word at the i-th position in the text, and N represents the number of words in the text; then the probability of each word in the text being replaced is calculated Where p is a hyperparameter, TFIDF(x i ) is the TFIDF value of the i-th word, max i TFIDF(x i ) is the maximum TFIDF value in sentence X; in each piece of text data, each word in the text is sampled according to its corresponding replacement probability T to determine whether the word is replaced with a TF-IDF non-core word; (2-3) Segment all the reverse-translated texts and build a replacement dictionary based on the segmentation results. The frequency of each word in the dictionary is For the words in the sentence that need to be replaced by TF-IDF non-core words, a word is extracted from the dictionary according to the probability p(w) for replacement, and finally the enhanced unlabeled text data is obtained, where each unlabeled text corresponds to an enhanced unlabeled text data; (3) Extracting BERT semantic and LIWC cognitive dual features of all discussion text data in the labeled dataset, unlabeled dataset, and enhanced unlabeled dataset; specifically including: (3-1) Extracting BERT semantic features: Input each discussion text in all data sets into the BERT model based on Chinese prediction pre-training, and select the first vector in the last layer output sequence as the semantic feature representation of the discussion text, that is: E bert ={e1,e2,…,e n }, where n is the dimension of the hidden layer in the BERT model; (3-2) Extracting LIWC cognitive features: Each discussion text in all data sets is input into the LIWC cognitive dictionary, and the LIWC cognitive feature E of each discussion text is calculated according to the following formula: LIWC ={e1,e2,…,e m } Among them, s is the sth LIWC cognitive feature, Dic s The corresponding dictionary; (4) Fuse the dual features of each discussion text to obtain new fused features; (5) Use labeled datasets with dual features, unlabeled datasets, and enhanced unlabeled datasets to train the cognitive engagement tracking model.

2. The cognitive engagement tracking method based on dual features and semi-supervised learning according to claim 1, characterized in that The specific steps of collecting, cleaning and labeling data in step (1) include: (1-1) Obtain the initial discussion text data through crawlers or backend data acquisition methods of online learning platforms; (1-2) Clean the HTML tags and irrelevant characters in the initial discussion text data by regular matching, simplify traditional Chinese characters, and remove stop words; (1-3) 10%-40% of all data are manually labeled.

3. The cognitive engagement tracking method based on dual features and semi-supervised learning according to claim 1, characterized in that The method of integrating the dual features of the discussion text in step (4) specifically includes: The BERT semantic feature E bert and LIWC cognitive characteristics E LIWC Linear splicing to generate an intermediate fusion feature E of length m+n BL ={e1,…,e n ,…,e m+n }, then the intermediate fusion feature is input into the linear connection layer according to the following formula, and then passes through the Tanh activation function to output the final fusion feature E con ={e1,e2,…,e k } E con =Tanh(W1·E BL +b1) Among them, W1 is the weight matrix of the linear connection layer, b1 represents the bias of the linear connection layer, and k is the output feature E after feature fusion. con Dimension.

4. The cognitive engagement tracking method based on dual features and semi-supervised learning according to claim 1, characterized in that The training process of the cognitive engagement tracking model for the text discussed in step (5) includes: (5-1) The fused features of all discussion texts are put into the fully connected layer respectively, and the output of the fully connected layer is followed by a softmax activation function to obtain the cognitive investment distribution corresponding to each discussion text; (5-2) Perform different calculations based on different data sets For a labeled data set, the loss function between the labeled data and the corresponding label is calculated according to the following formula: Among them, y is the label of the labeled data, p l is the cognitive investment distribution of the labeled data; For unlabeled data sets and enhanced unlabeled data sets, it is necessary to first sharpen the cognitive input distribution of the enhanced unlabeled data, and then calculate the loss function of the unlabeled data and the corresponding enhanced unlabeled data according to the following formula: Among them, p u is the cognitive input distribution of unlabeled data, is the sharpened cognitive input distribution of the enhanced unlabeled data; Finally, the two loss functions are used to calculate the objective function according to the following formula: Loss Final =Loss Supervised +λLoss Consistency Among them, λ is a hyperparameter used to control the impact of unlabeled data on model parameters; (5-3) The parameters of the cognitive input tracking model are updated by minimizing the objective function to complete the training of the cognitive input tracking model.

Images