Design and Application of English Language and Literature Smart Classroom Based on Artificial Intelligence Technology

Before class, teachers release pre-study materials in advance on the cloud teaching platform, arrange the learning content, and set up learning tasks; students organize their learning outcomes and record doubtful parties through independent learning and self-testing before class; teachers use the online pre-study data to provide good pre-assessment and analysis of the learning situation, optimize the design and curriculum, and guide the classroom teaching with the pre-study situation, so that the data-driven decision-making can really become a reality. .

In the smart classroom, the teacher, based on the data of students’ pre-course preparation and assessment results viewed and grasped in the background, carries out various forms of online and offline teacher-student interactions and student-student interactions through the teaching methods of task teaching method and guided inquiry method, guides students to actively participate in classroom teaching activities, mobilizes students’ enthusiasm for learning, and breaks the traditional teaching mode of “Teachers filling the classroom and students silently listening to the teaching”. The traditional teaching mode of “teachers filling the classroom and students listening silently” is broken. When students cooperate and explore, teachers can listen to and record their doubts, answer questions, and solve problems in a focused and targeted way. After students complete the exploration task, teachers can send the accompanying test to their smart terminals on the cloud teaching platform. Through the instant feedback data of the accompanying tests submitted by the students, they can fully grasp the learning situation of the students and realize “instant evaluation feedback”.

At the post-course stage, teachers check the background data and statistics of the teaching platform to understand the learning mastery of each student before and during the lesson, assign personalized hierarchical assignments to students according to the differences in their learning conditions, provide differentiated and personalized tutoring and Q&A, and focus on guiding the students to learn independently and consolidate what they have learned. In addition to traditional written assignments, students’ after-school assignments are more diverse, such as online tests, audio or video assignments, and so on. Students can submit their after-school assignments automatically. The post-course assignments submitted by students can be automatically reviewed by the system to receive instant feedback. The English Language and Literature Smart Classroom Teaching Model records students’ learning dynamics in an all-round way through the background data of the teaching platform, assisting teachers to grasp the learning data in real time and pay more attention to the students’ learning process.

Long Short-Term Memory Network LSTM is improved on the basis of Recurrent Neural Network RNN, utilizing the gating mechanism to update the cellular state and the hidden state, which solves the problem of long-time dependence in RNN, and is widely used in the fields of speech processing, data analysis, and prediction. The most representative of the long and short-term memory network LSTM is its three gates, namely the forget gate, input gate and output gate, which have their own different roles, and can effectively alleviate the phenomenon of overfitting and underfitting in the training process through the gating mechanism.

C_t–1 represents the cell state of the previous LSTM cell, C_t is the cell state of the LSTM cell in the current state, and Ct′ is the intermediate state. h_t–1 is the hidden state of the previous LSTM cell, h_t is the hidden state of the LSTM in the current state, i.e., the output value of this cell, and x_t is the input value at this moment in time.The LSTM network is inputting h_t–1 & x_t into the forget gate f_i, which serves to forget unimportant information. The exact formula is shown in (1): 1 ft=σ(w[ ht−1,xt ]+bf) i_t is an input gate that is capable of selectively memorizing the information and putting it into the cell state C_t as shown in (2): 2 it=σ(w[ ht−1,xt ]+bi) C_t is specifically calculated by dot-multiplying with the input gate information, which in turn updates C_t. It is responsible for memorizing the information that needs to be memorized within each cell at the current moment and discarding the information that does not need to be memorized in the previous cell before finally passing it to the next cell. The specific computational formulas are shown in (3) to (4): 3 Ct′=tanh(w[ ht−1,xt ]) 4 Ct=ft⋅Ct−1+it⋅Ct o_t is the output gate, which is able to decide how much information in the current state needs to be output, as shown in (5): 5 ot=σ(w[ ht−1,xt ]+bo) h_t is the hidden state also the final output, which is calculated as shown in (6): 6 ht=ot⋅tanh(Ct)

In the above formula, σ represents the sigmoid function, which is able to map the data between [0,1], and the tanh function is able to output the data between [-1,1] and centered on 0, which can prompt the model to undergo faster convergence, while w and b are the weights and biases obtained through training. During the development and application of deep recurrent neural networks, the processing of long sequences has always been a difficult problem, and there is always a long time dependence problem, which is effectively solved in LSTM networks. Compared with RNN, LSTM has three gating units and corresponding hidden states, which can control the input and output of information, and can better deal with the gradient disappearance and gradient explosion problem in the training process, so LSTM has a wider range of applications.

In the previous paper, the LSTM memory neural network model is introduced to have the best generalization in the field of natural language processing, so the study chooses the long and short-term memory neural network model as the basis for constructing the English language and literature appreciation model. Compared with other deep neural network models directly input works for training, in this paper, we do not directly input articles into the neural network for training, in this paper, we will choose the sample set (scoring rules) as part of the input, the sample set to choose different score samples, calculate the textual differences between the articles, indicating the differences in language features, so as to represent the “scoring rules “.

The model consists of three modules, Ma, Mb, and Mc, with different combinations of modules receiving different inputs. The combination of modules Ma and Mc receives only articles, Mb and Mo receive distance information, and Ma, Mb, and Mc receive both articles and samples.

The Embedding Layer is represented as a fixed-length word vector, in which all word vectors are populated to their maximum length in this sequence. Then, each word vector is converted into a sequence of low-dimensional vectors through the embedding layer. For ease of description, function v is used in the study to represent the word embedding process. v(e_i)∈R|V|×D and v(ej)∈R^|V|×D are the word embedding outputs, where |V| is the size of the word and D is the dimension of the word embedding. After word embedding, dist_i,j =v(ei)–v(sj) is used to represent the distance information between articles v(ei) and v(sj).

Convolution Layer. This layer is not required and can be skipped, particularly for shorter articles. For longer articles, it may be helpful for the network to extract local features from the word vectors before applying the recurrent neural network. This optional feature is implemented by adding a convolutional layer to the embedding layer.

LSTM network layer. The word embedding sequence obtained from the embedding (or convolutional) layer is passed to the long short-term memory network to compute Equation (7): 7 ht=LSTM(ht−1,xt) where xt and ht are the input vectors at moment t. The LSTM model controls the flow of information in the recursive operation by parameterizing the output, input and forgetting gates. The following equations formally describe the LSTM function: 8 it=σ(Wi⋅xt+Ui⋅ht−1+bi) 9 ft=σ(Wf⋅xt+Uf⋅ht−1+bf) 10 ct=tanh(Wc⋅xt+Uc⋅ht−1+bc) 11 ct=it∘ct+ft∘ct−1 12 ot=σ(Wo⋅xt+Uo⋅ht−1+bo) 13 ht=ot∘tanh(ct)

At moment t, the LSTM outputs a hidden vector ht, ht reflecting the semantic content of the article at t. In the experiments, two LSTMs and the attentional mechanism in the LSTM layer were chosen to be used.

Self-characterization layer. Let he be the implicit layer of the article and het be his vector at position t, and let hd be the distance information implicit layer and hdt be the vector of hd at position t. Let d be the sentence length (assuming that different sentences have the same length). Then, the similarity of the vectors at positions t and t+d is calculated, which is called intrasimilarity feature in the study, and the formula (14) is calculated as follows: 14 inner−feature=het⋅ het+δ∣het|⋅|het+δ∣

In addition, the similarity between vectors he and hd at the same position t is calculated, which is called the crossover feature in the study, and is shown in Equation (15): 15 cross−feature=het⋅het∣het|⋅|het∣

The internal features and cross features are connected to vectors and output to the next layer, respectively. In addition to the inner and cross features, this layer has two main outputs: the thesis-implicit layer and the distance-informative-implicit layer. Two kinds of processing are performed on these two layers. One method is to directly take the vectors of the last positions of he and hd. The other method is to take the vector of mean values over time. Name these two vectors he–vector and hd–vector. Four vectors are fed into the fully connected layer.

Full Link Layer. Get four vectors from its own feature layer: he–vector, hd–vector , inner–feature and cross–feature. Link these four vectors into one vector, input vector to Softmax layer.

Softmax layer. This layer classifies the output of the fully connected layer. It classifies the output by using the equation, which is calculated as shown in (16): 16 s(x)=sigmoid(w⋅x+b) where x is the input vector (the output of the fully connected layer), w is the weight vector, and b is the bias.

The dataset used in this paper is from an automated scoring program (ASAP) for essays from the kaggle 2012 competition. In total, the dataset collects essays from more than 17,000 middle school students in the U.S. and provides the results of manually scoring the essays.The dataset consists of eight categories, which are divided into the following proportions: 60% of the data is the training set, 20% of the data is the validation set, and 20% of the data is the test set.

In this paper, consistency metrics are used to assess the degree of consistency between the scoring model and the manual scoring. This is also the official evaluation method used by the ASAP program mentioned above. Many consistency metrics are applicable to this automatic scoring model for essays, such as Pearson’s correlation, Spearman’s rank correlation, simple Kappa coefficient, quadratic weighted Kappa coefficient, and so on. In this paper, we use the quadratic weighted Kappa (QWK) coefficient, a consistency indicator, which ranges from 0 to 1, with larger values indicating a higher level of consistency, 0 indicating basic inconsistency, and 1 indicating basic consistency.

With the rapid development of NLP technology and deep learning, as well as the widespread application of essay scoring systems, more and more people begin to rely on automatic essay scoring methods based on deep learning. LSTM-MoT, CNN-CNN-MoT, and CNN-LSTM-ATT are selected as reference objects for the experiments in this paper.

The above proposed improved two-layer LSTM based on LSTM is used to automatically score English language and literature, and the whole neural network from semantics (LSTM_sem), semantics and shallow language features (LSTM_sem+simple), semantics and syntax (LSTM_sem+synt), semantics and topic relevance (LSTM_sem+topic), and the whole neural network (LSTM_All) are separately experimented on eight essay collections from the publicly available dataset ASAP, and the results are compared to several of the baseline models described above.

In this section, a comparative analysis of different models on Quadratically Weighted Kappa Coefficient (QWK) will be conducted for each of the eight essay collections based on the degree of consistency between model scoring and manual scoring. In addition, this paper also conducts related experiments using the improved two-layer LSTM model to investigate the effect of different aspects of textual features on the scoring results of the works.

The experimental comparisons of different models on QWK assessment metrics are shown in Table 1. The average QWK of the improved two-layer LSTM model for English language and literature works appreciation proposed in this paper is 0.810, which is significantly better than other baseline methods, and in the assessment metric of average quadratic weighted kappa coefficient this paper’s model improves by 9.76% compared to the baseline models LSTM-MoT, CNN-CNN-MoT, and CNN-LSTM-ATT, respectively, 7.14% and 3.18% respectively.

The comparison results of the models on different text features regarding QWK are shown in Table 2. The models combining semantic and other textual features (LSTM_sem+simple, LSTM_sem+synt, LSTM_sem+topic) all improve compared to LSTM_sem alone, and the improved two-layer LSTM model achieves the best experimental results in terms of semantic and shallow linguistic features, i.e., LSTM_sem+simple. The specifics are as follows:

On the average quadratic weighted Kappa coefficient assessment index, LSTMsem+simple reaches 0.835, which is 9.72%, 6.91%, 5.16%, and 3.09% higher than the LSTM_sem+synt model, LSTM_sem+topic model, and LSTMAll model, respectively. This indicates that compared to syntactic features, shallow linguistic features have a greater impact on the scoring results of English language literature. However, compared to the LSTM_sem+topic model that combines semantic and topic relevance, the scoring results of the semantic and syntactic based LSTM_sem+synt model have a higher degree of consistency with the manual scoring results, which indicates that in addition to the content features, the linguistic features of the text are also an important factor that needs to be considered in scoring the works.

To confirm whether there was a significant difference between the two classes before the implementation of the instruction, descriptive statistics were conducted on the two classes using SPSS. The results of the pre-laboratory test in English Language Arts are shown in Table 3.The mean values of Class 1 and Class 2 were 42.56 and 43.07, and the difference between the mean scores was 0.51, with the mean value of Class 2 being slightly higher than that of Class 1.The standard deviation of Class 1, 5.712, was higher than that of Class 2, 5.324, and the data suggests that the degree of dispersion in the achievement of English Language Arts among the students in Class 1 is high compared to that of the students in Class 2.Sig. = 0.286 >0.05, so the hypothesis that the two classes have equal variance on the achievement variables cannot be rejected, and the equal variance of the pre-test achievement data is also a direct indication that the two classes are at a comparable level of achievement in English language and literature. To sum up class 2’s average grade is slightly higher than class 1, but analyzed by independent samples t-test it is concluded that there is no significant difference in English language and literature grades between the two classes, and the level of English language and literature is roughly comparable, which is suitable for the object of this study.

In order to confirm whether the smart classroom teaching model has an improving effect on students’ English language and literature competence in teaching English language and literature, the post-test scores of English language and literature in two classes were analyzed after the experiment.

The results of the English Language and Literature post-experimental test are shown in Table 4.The post-test scores of Class 1 and Class 2 were 45.82 and 43.54 respectively, and the mean score of Class 1 was significantly higher than the mean score of Class 2 after the implementation of Smart Classroom Teaching. And the standard deviation of class 1 scores is 3.862, much smaller than the standard deviation of the control class of 5.238, indicating that the degree of dispersion of students’ scores on this test in class 1 is lower than that in class 2, and students’ scores in each grade on the posttest are closer to the mean, and the overall mean scores have improved.

The correlation coefficient between the English Language and Literature pre-test scores and post-test scores of the students in Class 1 is 0.528, with a significant level value of 0.001, which proves that there is a linear correlation between the pre-test and post-test scores of English Language and Literature proficiency of the students in Class 1, after they were influenced by the smart classroom teaching model, on the basis of this level of significance.

The results of the paired samples t-test show that the difference between the mean scores of the pre and post-tests of Class 1 is 3.260, the standard deviation of the difference and the standard error of the mean are 5.315 and 1.032 respectively, and the t-value of the pre and post-test data is -2.175, with the degree of freedom df=46, so the significance probability of the two-tailed test is p=0.022, which is smaller than α=0.05, so it indicates that Class 1 has a more significant difference in the smart classroom teaching mode of teaching there is a more significant difference between the total achievement scores of English Language Arts in the pre and post-test, and there is a small increase in the total achievement scores of English Language Arts. However, the data also shows that the standard deviation is large, which indicates that the gap between the scores and the mean score is large. To explore the changes in the distribution of the scores, the pre- and post-test reading scores were segmented into steps of 5.

The distribution of test scores in English Language and Literature is shown in Figure 4.The scores in the pre-test of Class 1 are mainly concentrated in the range of 40 to 45, and the number of students in the high score band of over 50 is relatively small. In the post-test scores, the score range is concentrated in the range of 40 to 50, and the number of students in the high score range of 50 or more has increased significantly, from 23.4% to 31.9%, and the number of students in the middle and high score ranges has also increased significantly, which also indicates that there is an overall level of improvement in the English Language and Literature scores of the class, especially the students with medium and above English proficiency levels are more obviously affected by the role of the smart classroom.

Figure 4.

The distribution of test scores in English language and literature

English Language and Literature competence was divided into four aspects: summarizing the main idea, finding detailed facts, guessing the meaning of words, and reasoning and judging, and English Language and Literature test questions from the pre and post-tests of the Class 1 instructional implementation were divided according to the four aspects.

Statistical analyses of the corresponding question types of the four aspects of English Language and Literature comprehension on the Class 1 pre and post-tests were conducted to investigate the effect of the smart classroom on the different aspects of English Language and Literature proficiency. The results of the statistical analysis of the results of each question type are shown in Table 5. The mean scores of summarizing the main idea, finding detailed facts, guessing the meaning of words, and reasoning judgment have all been improved, and the standard deviations of all four question types have been reduced to a certain extent, which means that the scores of all four aspects have been improved in terms of the degree of dispersion compared with the mean scores of the pre-test.

The p-values for summarizing main idea and finding detailed facts are 0.002 and 0.006, and under the two-tailed test criterion (α=0.05), there is a significant difference between the pre and post-test scores of the two question types, and in view of the limited class scores, it can be judged that there is a tendency to improve the level of the English language and literature of the question types’ counterparts of summarizing main idea and finding detailed facts under the influence of the smart classroom. The p-values of guessing word meaning and reasoning judgment are 0.073 and 0.081, and the p-values are greater than the critical value under the significance test (α=0.05), so it can be judged that the students in class 1 do not have significant differences in the scores of the pre and post-tests although the mean scores of these two question types have slightly improved, indicating that there is no significant change in the scores of guessing word meaning and reasoning judgment under the influence of the smart classroom.