Online English writing teaching method that enhances teacher–student interaction

3.1 Improved neural matrix decomposition (INMD) construction based on knowledge graph enhancement

An online English writing-assisted teaching model for personalized exercise recommendation is created by using students’ online assessment exercise data to enable students to choose from a set of questions the exercises that match their present level of study. The exercise recommendation model is made up of an Exercise Knowledge Graph structure and an INMD model. The creation of the Exercise Knowledge Graph is shown in Figure 1.

Figure 1

Schematic diagram of the establishment process of exercise knowledge map.

As can be observed from Figure 1, the exercise knowledge map first establishes the exercise nodes, then establishes the exercise and knowledge point relationships, and finally establishes the exercise and event relationships [15]. The exercises are constituted by a number of knowledge points. Knowledge points represent the smallest units of relatively independent information such as knowledge and theory. Events represent the set of exercises that students need to complete within a limited time frame.

The exercises will be designated as T = ( k 1 , k 2 , … k n ) , the events as C = ( t 1 , t 2 , … t n ) , and the knowledge points as k . The exercises’ knowledge base will be referred to as G = { ( h , r , t ) ∣ h , t ∈ E , r ∈ R } . Exercises, knowledge points, and events are all included in the collection of entities known as E . R stands for the logical connections between entities [16]. Examples are denoted by h and t . Exercises serve as test questions with reference answers that students can understand. Exercises, knowledge points, and events are all examples of the entity set, as can be observed, and when these two are translated into the knowledge graph, an exercise node, a knowledge point node, and an event node will arise, respectively, adding a relationship between the two nodes. Figure 2 shows the INMD exercise recommendation model.

Figure 2

INMD exercise recommendation model.

A feature sharing unit, a feature improvement unit, a neural matrix decomposition unit, and a knowledge graph embedding unit make up the model, as shown in Figure 2. The enhanced neuromatrix decomposition exercise recommendation model has a multi-task learning architecture and is trained in an alternating learning mode, enabling the alternate training of the exercise recommendation task and the knowledge graph learning job [17,18]. The knowledge graph feature learning task is integrated into the training process of the recommendation task by the feature sharing unit through crossover and compression operations, enhancing the effectiveness of the recommendation task with the knowledge graph data. The feature sharing unit’s structure is shown in Figure 3.

Figure 3

Structural diagram of feature sharing unit.

A vector input layer, crossover operation processing, crossover feature matrix creation, compression operation processing, and a vector output layer are all components of the feature sharing unit, as shown in Figure 3. The vector compression operation is carried out by the unit after the vector crossover operation. The two one-dimensional input vectors are ready to be operated upon at the crossover step. The operation is a vector multiplication, and the resulting two-dimensional matrix is then used to merge the data [19]. The weight matrix is multiplied by the two 2D matrices containing the input vector information during the vector compression step to produce the compressed 1D vectors for sharing vector information. It is clear that the feature sharing unit will be crucial to raising the effectiveness of the recommendation model. The crossover operation’s mathematical equation is as follows:

where I l and H l stand for the input one-dimensional vectors. The vector length is denoted by d . C l stands for the two-dimensional cross-featured matrix. The compression operation’s mathematical expression is as follows:

The weight matrix and bias vector necessary for the compression process are denoted in equation (2) by the letters w l and b l , respectively. The new vector after compression is indicated by I l + 1 and H l + 1 , respectively. A feature enhancement unit receives the compressed new feature vector input. Factorization machine (FM) and deep neural networks (DNN) are used by this unit to capture the vector’s lower-order features and higher-order features, respectively. The mathematical equation for feature improvement is as follows:

where x i stands for the numerous attribute data used in student learning. The weight coefficients are denoted by α i . The sigmoid function is represented by σ . M stands for the neural network that is fully connected. The feature vector after integrating the features from FM and DNN is represented by the letter U L . The neural matrix decomposition unit can calculate the learner’s preferred exercise by entering U L , as shown in the following equation:

where the letters × , p U G , and p U M stand for vector multiplication, generalized matrix decomposition, and multilayer perceptron, respectively. y ˆ U I indicates how much the learner liked the exercise. The knowledge mapping embedding unit’s mathematical calculation is expressed in the following equation:

where t ˆ stands for the expected value of the vector corresponding to the lower node of knowledge. For the purpose of acquiring t ˆ , this unit will take the features from the knowledge superordinate phases and relationships and combine them. The phrase for determining how similar the expected and actual values are is found in the following equation:

where f K G denotes the similarity calculation function and ( H , R , t ) denotes the knowledge triad.

3.2 Interactive short-text sentiment analysis model construction

The study began with textual sentiment analysis to look at students’ attitudinal tendencies toward learning comments in order to objectively analyze students’ interactive qualities in online learning. The brief text type includes the interactive remarks made by students. Due to the considerable feature sparsity of this text type, direct text sentiment analysis typically produces inaccurate results. To improve the extraction of sentiment data from short messages, a text extension method will be employed. The approach begins by pre-processing data for short-text datasets. To increase data purity, pre-processing cleans and eliminates abrasive information from the raw data, such as stop words and pointless symbols. The pre-processed text data type must be converted into numerical input data based on a vectorized representation because it cannot be supplied directly into the algorithm model. The Word2Vec model often serves as the foundation for the vectorized representation of text. The Continuous Jumping Word Model and the Continuity Bag of Words (CBOW) model make up the model. The CBOW model was chosen because of its quick transformation of text properties into high-dimensional vectors, which was all that the investigation required. The CBOW model’s structure is shown in Figure 4.

Figure 4

Structural diagram of the CBOW model.

As seen in Figure 4, the CBOW model’s input layer uses one-hot coding to convert the text into an N-dimensional vector. The input layer contains n N-dimensional vectors. The letters W stand for the mapping weight matrix, N stands for the number of words in a phrase, and M stands for the dimensionality of the word vector. The N-dimensional word vectors are averaged to get the implied layer vectors. A classifier with Softmax is used as the output layer, and the output values are probabilities. For the calculation of sentiment relevance, the vectorized representation of the text features will be employed. Lexical classification weights are calculated before the sentiment relevance computation. The study uses Word Frequency-Inverse Document Frequency (TF-IDF) to calculate the importance of different words and obtain their weights for classification purposes. The TF-IDF calculation is expressed as:

where n i , d represents the quantity of instances of feature f in document d , ∑ k n k , d represents the total number of times each feature appears in document d , and d indicates how many articles include feature f i . The uniformization of word frequencies in equation (8) prevents the detrimental effects of significant gaps in text length:

The unimproved TF-IDF does not account for feature category information, which can reduce the algorithm’s clustering power for short texts. As a result, the study will use the term “information gain” to define the degree to which features contribute to the themes that are categorized. The information gain is expressed in the following equation:

where X denotes the text set, Y denotes the text features, H ( X ) denotes the entropy value of the text set to describe the uncertainty of the classification of text X , and H ( X , Y ) denotes the uncertainty of feature Y for the classification. When the information gain of all features is calculated, comparing the size will obtain the maximum value of information gain F i g max . At this point, if k in data set is obtained by random sampling from the original data set and F − { F i g max } is used as the feature set, k decision trees can be constructed. Let k and N i j denote the number of positive samples for classification before and after adding noise to the features, respectively, then the quantified value of the thematic relevance of a feature in each tree is I i j = ∣ M i j − N i j ∣ . Thus, the sentiment relevance can be calculated based on the following equation:

where 1 k ∑ 1 k I k i stands for the median topic relevance, while R ( f i ) stands for the median sentiment relevance. By ranking all sentiment relatedness in descending order, the top word vectors were chosen to create the keyword set. The density of the keyword set was then increased by applying an expansion method based on autocorrelation association. An illustration of a self-associative word association network may be found in Figure 5.

Figure 5

Example diagram of self-associative word association network.

The word association network is made up of three components, as shown in Figure 5: graph nodes, connected edges, and weight values. The weights are used to gauge how closely subsets are related. Co-occurrence probabilities and feature dependencies are computed to construct a word association network. The co-occurrence probability shows the likelihood that one trait will emerge when another one does. The percentage of the feature set’s feature sequence that is represented by feature dependency. The co-occurrence probability can be calculated as follows:

where both f i and f j denote the features, P ( f i , f j ) denotes the probability of both features occurring at the same time, and P ( f i ) denotes the probability of feature f occurring. The calculation of feature dependence is expressed as follows:

where freq represents the feature frequency The I ( f i , f j ) indicates how closely traits f i and feature f j are related.

4.1 Utility analysis of INMD models based on knowledge graph enhancement

Exercises, competitions, user data, user records, and knowledge data will all be included in the dataset for the study, which was derived from the online testing platform Codeforces. There are 6,920, 1,315, 45,469, 89,365, and 262 entries in total. The event name and the event exercise set make up the event data. The submission result, event name, and submission time are all included in the user record, as shown in Table 1.

In this study, the training and test sets were split into two groups in a 3:1 ratio. Accuracy, recall, F1 value, and diversity mean were chosen as the measures for evaluating the model. The knowledge graph’s vector length was also set to 8, the knowledge graph embedding model’s learning rate was set to 0.02, and the INMD exercise recommendation model’s learning rate was set to 0.05. The recall rate and F1 value comparison curves for several models are shown in Figure 6.

Figure 6

Comparison curve of recall rate and F1 value of different models: (a) recall and (b) F1 value.

As can be noted from Figure 6, the proposed model obtained the highest values in terms of recall as well as F1 values. The proposed model achieved the highest recall and F1 values. The collaborative filtering algorithm (CFA) and deep factor decomposition module (DeepFM) observed the second highest and lowest values, respectively. All three models’ recall and F1 curves exhibit a sharp ascent followed by a gradual turn toward convergence. The recall measure for the proposed model starts to converge at a value of 84.6%, whereas the corresponding values for the CFA and DeepFM methods are 70.1 and 58.2%. When there had been 200 iterations, the recall of the suggested model had finally converged to 94.6%, as opposed to 90.2 and 88.5% for the other two algorithms. At the end of the iterations, the recall metrics of the proposed model outperformed those of the CFA and DeepFM algorithms by 4.4 and 6.1%, respectively. Additionally, the proposed model’s ultimate convergence rate was 92.3% when 200 iterations were used, as opposed to 90.0 and 88.2% for the CFA and DeepFM methods. As can be observed from the comparison, the suggested model improved by 2.3 and 4.1%, respectively. This demonstrates that the suggested model may successfully include the knowledge graph information from the exercises and enhance the efficacy of the task of providing learners with personalized exercise recommendations. Figure 7 displays the mean diversity curves of the comparison models as well as the mean diversity curves for various matching criteria.

Figure 7

Diversity mean curve under different matching conditions and comparison model diversity mean curve: (a) mean value of diversity under different matching conditions and (b) comparison model diversity mean.

Figure 7 illustrates how the study’s matching conditions were Pass, Interest, Suit, and Interest with Suit (IWS). For each of the several matching criteria, the study’s suggested mean values for model diversity varied. When there were 20 iterations, IWS increased the diversity mean metric over the first three matching conditions by 0.19, 0.16, and 0.09, respectively. When there were 120 iterations, or when the iterations were finished, the IWS matching condition had the highest diversity mean (0.83), compared to the Pass, Interest, and Suit matching conditions’ 0.74, 0.55, and 0.48, respectively. The comparison reveals that the IWS matching condition is better than the subsequent three by 0.09, 0.28, and 0.35. The diversity mean curves for the three various models under the IWS matching requirement are displayed in Figure 7(b). The experiments show that the mean value of diversity for the proposed model is 0.93 at the end of the iteration, while the corresponding mean values for CFA and DeepFM are 0.79 and 0.70, respectively. The comparison reveals that the proposed model outperforms the other two algorithms by 0.14 and 0.23. Therefore, this indicates that the model can simultaneously consider the learner’s interest and difficulty adaptation in the exercises, making the diversity of exercise recommendation results the most abundant. The accuracy and loss value comparison curves for the various models are displayed in Figure 8.

Figure 8

Comparison curve of accuracy and loss value of different models: (a) DeepFM, (b) CFA, and (c) model in this article.

The loss value curve for the DeepFM algorithm’s test set declines at the slowest pace, followed by the loss value curve for the proposed model’s test set, which decreases at the fastest rate, and the loss value curve for the CFA method, which decreases at a moderate rate. Before 200 iterations, the loss value of the CFA algorithm’s test set began to converge, eventually settling at 23.5%. After 150 iterations, the loss value curve for the test set of the suggested model begins to converge, with a final convergence value of 15.8%. The proposed model achieved the greatest value of recommendation accuracy based on the test set. Figure 8(c) shows that the model’s accuracy curve climbs more quickly and begins to converge at iteration number 170, eventually reaching a maximum accuracy of 88.6% at iteration number 250. The most accurate algorithms, CFA and DeepFM, had accuracy rates of 78.2 and 68.9%, respectively. The comparison reveals that the proposed model outperformed the latter two algorithms in terms of accuracy metrics by 10.4 and 19.7%. This proves that the model is better at recommending exercises that are specific to each individual.

4.2 Analysis of the utility of the sentiment analysis model for interactive short texts

Student interactive texts from the MOOC website’s College English Writing course were chosen for analysis as part of the study’s effort to evaluate the efficacy of the sentiment analysis model of interactive texts based on sentiment relevance. Positive and negative attitudes were separated into separate categories in the student interactive text. Three dimensions made up positive emotions: satisfaction, happiness, and enjoyment. There were three types of negative emotions: dissatisfaction, dislike, and boredom. Accuracy, recall, and F1 value were chosen as the model evaluation criteria. Random Forest and SVM were the comparison algorithms of choice. The accuracy curves for the three models are shown in Figure 9 both before and after the brief text expansion.

Figure 9

Accuracy curves of three models before and after short text expansion: (a) SVM, (b) Random Forest, and (c) model in this article.

Figure 9 shows that, regardless of the model, the accuracy curves before and after the expansion grew. When there were 100 iterations, the pre-expansion accuracy convergence value for the SVM model was 69.3%. The accuracy convergence after extension increased by 14.1% between before and after, to 83.4%. At the end of the iterations, the Random Forest model’s pre- and post-expansion accuracy convergence values were 83.8 and 87.6%, respectively. Pre- and post-extension accuracy for the study’s model was 82.1 and 89.3%, respectively, indicating an improvement in accuracy of 7.2%. The comparison revealed that the study’s model outperformed the first two models after extension by 5.9 and 1.7%. Figure 10 displays the F1 value curves and recall for the three models both before and after the brief text expansion.

Figure 10

Recall rate and F1 value curve of three models after short-text expansion: (a) recall and (b) F1 value.

Figure 10(a) illustrates the overall growing trend in recall curves for the extended SVM, Random Forest, and the study’s adopted sentiment relevance-based interactive text sentiment analysis model. The SVM model, which had a size of 88.1%, had the highest recall rate among them after 90 iterations. With a size of 84.5%, the Random Forest model achieved the highest recall after 50 iterations. The study’s model, which had a size of 94.3%, had the highest recall after 90 iterations. The comparison reveals that the model’s recall is up to 6.2 and 9.8% from the previous two models. The F1 value curves of the extended SVM and randomness forest models all exhibit a rising and then a declining trend, while the F1 value curves of the study’s suggested model exhibit an increasing trend, as can be shown in Figure 10(b). The highest F1 value of the SVM model among them is 87.9%. The Random Forest model’s greatest F1 value was 88.3%. The study’s model has an F1 value as high as 90.4%. Comparing the two models reveals that this model has boosted the F1 value by 2.5 and 2.1%. This indicates that the model has an advantage in interactive text sentiment analysis. Figure 11 shows the results of the performance metrics of the model proposed in the study under different extensions.

Figure 11

Performance index results of the model proposed in the study under different expansion methods: (a) training set and (b) test set.

Figure 11 shows that in terms of accuracy, recall, and F1 value for both the training and test sets, the associative word association expansion technique used in the study is superior to the synonym expansion approach and the external corpus expansion approach. The study’s expansion method yields accuracy rates of 90.1%, recall rates of 96.8%, and F1 values of 96.8% for the test set, which are 9.6, 10.5, and 8.9% higher than the synonym expansion method and 6.7, 7.2, and 9.2% higher than the external corpus expansion method, respectively.