Research on the Innovative Model of English Teaching by Integrating Traditional Culture and Artificial Intelligence

Given a respondent’s answer sequence X=({q1,y1},{q2,y2},⋯,{qt−1,yt−1}) , where q_t is the question number and y_t ∈ {0, 1} indicates whether the respondent answered the question correctly or not, with correct being 1 and incorrect being 0. The main task of knowledge tracking is to predict the probability that the respondent will answer a new question correctly p(yt=1|qt,X) in time t, based on the given answer sequence X.

The desired features are extracted based on the sequence X: 1)

Repetition Time Gap (RTG), i.e., the delay between a student’s answering interaction for an exercise in the same knowledge point and the previous interaction.

Based on the overall architecture of DKVMN, the features of repetition interval and past answer times are added to its prediction part, and then the gating mechanism is utilized for feature fusion, and the improved DKVMN-F network structure is shown in Fig. 1. The DKVMN-F framework consists of embedding layer, key processing layer, reading layer, writing layer, and prediction layer.

Figure 1.

DKVMN-F Model Framework

The implementation process of DKVMN-F model is as follows: 1)

Initialization of mastery level. In DKVMN, the key matrix stores the potential knowledge points and the value matrix stores the mastery of knowledge points. Set the dimensions of key matrix (size N × d_k) and value matrix (size N × d_v), where N is the number of potential knowledge points, d_k, d_v are the dimensions of embedding, and randomly initialize the key matrix and value matrix.

That is, its correlation with each potential knowledge point. 4)

Knowledge point mastery calculation. The knowledge point mastery required for exercise number q_t is calculated at the reading layer based on the weights w_t obtained from the correlation calculation, i.e., the weighted sum of the inner product of w_t and each cell in the value matrix Mv=(Mv(1),Mv(2),⋯,Mv(N)) , the required knowledge point mastery: (2)rt=∑t=1Nwt(i)Mtv(i)

The residual linkage method was used to combine the exercise number embedding k_t with the knowledge point mastery r_t and to introduce the learning factor vector l_t to obtain a vector of summarized mastered information: (3)ot=Tanh(W1T[rt;kt;lt]+b1)

Where W₁, b₁ are the weight matrix and bias vector of the mastered knowledge respectively, so as to improve the performance of prediction. 5)

Knowledge mastery update, the update operation is performed on the value matrix Mtv at the write layer, i.e., based on the question q_t and the answer status y_t to q_t, the knowledge mastery of the student is obtained. For the embedding vector v_t, the forgetting information vector e_t is computed first: (4)et=Sigmoid(WeTvt+be)

where W_e, b_e are the weight matrix and bias vector of the forgotten information, respectively.

Next, the value of each knowledge point in the value matrix after forgetting is calculated: (5)M˜t+1′(i)=M′t(i)(1−wt′(i)et)

Where wt′(i) is the correlation weight with each knowledge point obtained using correlation calculation.

Then the vector of added information learned from the topic is calculated a_t: (6)at=Tanh(WaT⋅vt+ba)

where W_a, b_a are the weight matrix and bias vector of the added learning information, respectively.

Finally the updated memory matrix is: (7)Mt+1(i)=M˜t+1v(i)+wt′(i)at

6)

Answer prediction, using the forgetting gate mechanism to integrate the degree of mastery of knowledge points to obtain the total vector of learning information o_t and the total vector of forgetting: (8)fst=wt⊙ft

where f_t is the forgetting factor vector.

First, the gate vector g_t ∈ R^N is computed for controlling the reading and writing of memories: (9)gt=Sigmoid(Wo⋅ot+Wf⋅fst)

where W_o ∈ R^N×N and W_f ∈ R^N×N are the weight matrices of total learning information and total forgetting information, respectively.

Next, the combined information vector glt∈RN integrating forgetting and learning is obtained: (10)glt=gt⊙ot+(1−gt)⊙fst

Finally, the combined information vector glt is sent to the Sigmoid layer to calculate the probability of correctly answering the current exercise: (11)pt=Sigmoid(WgT⋅glt+bg)

where W_g ∈ R^N, b_g are the weight matrix and bias vector of the combined information. 7)

Model optimization. The model is optimized using the cross-entropy loss function between the predicted exercise answer status and the real answer status: (12)L=−∑t(rtlogpt+(1−rt)log(1−pt))

For any two topics E_a and E_b, this paper uses S(Ea,Eb) to evaluate the similarity between them. The higher the score of S(Ea,Eb) , the higher the similarity between E_a and E_b. In the absence of loss values, the Recommended Similar English Topics (RSE) can be formulated as: (13)F(E,R,Θ)→Rs

Θ is the parameter of F, R=(E1,E2,E3,⋯) is the set of similar topics for topic E, and Rs=(E1s,E2s,E3s,⋯) is derived in descending order according to the similarity score s(E,E1s),s(E,E2s),⋯ , for E the higher the similarity, the higher the score he gets.

In order to solve the above problem, this paper proposes a two-phase solution that includes a training phase and a testing phase. In the training phase, the text of the topic, the knowledge points of the topic and the source code of the topic are given. In this paper, we propose the HANN framework to learn a unified semantic representation of each topic through multi-modal processing of heterogeneous and structured data, while calculating the similarity scores of each team of topics. Then a pairwise loss function is utilized to train the HANN. After obtaining the training results of the HANN, for any topic E_a which is in the testing phase, its similar topic (Ea,1s,Ea,2s,⋯) can be found and the set of similar topics is sorted by ranking them based on the similarity score.

The specific flow of the Heterogeneous Data Attention Neural Network (HANN) framework is shown in Fig. 2. HANN consists of three main parts: the Heterogeneous Data Representation Layer (HERL) of the topic, the Attention Layer (SA), and the Prediction Scoring Layer (SL).

Figure 2.

Flowchart of the HANN framework

1)

HERL layer

For the input of HERL, firstly, the topic text features are extracted, the topic knowledge points are transformed into a knowledge point matrix (Q matrix), and the topic codes are one-hot coded to extract the code features. Then, an Attention-based Long and Short-Term Memory Network (Attention-based LSTM) is designed to learn a unified semantic representation of each topic by integrating different topic data through LSTM [27]. (1)

Topic Input: for the HERL framework the input is the data of the topic E, e.g., the text (ET), the topic source code (EC) and the topic knowledge matrix (EQ), which will be (ET) formalized as N a sequence of pairs of words ET=(w1,w2,⋯wN) where wi∈ℝd0 , the word2vec initialized by a d₀ trained word embedding. for the topic code, the keyword information of the code is extracted, and then encoded in a one-hot encoding, the code in the topic E can be represented by A matrix representation EC=(k1,k2,⋯kL)∈{0,1}L×Latt where k_i is a one-hot vector whose dimension is equal to the total number of keywords in all the codes in the topic library, and L_att is the number of keywords of the codes in the topic E. The knowledge points of the topic will be represented as Q matrix information EQ=(q1,q2,⋯qL)∈{0,1}M×Natt .

where Wu∈ℝLatt×d2 is the parameter of the embedding layer and ui∈ℝd2 is the output parameter of the embedding layer. Thereby, the source code keyword EC is transformed into a matrix u=(u1,u2,⋯uL)∈ℝL×d2 .

(3)

Attention based LSTM after obtaining the feature representations of the knowledge points and the code, by integrating all the heterogeneous data entered into the topic E, i.e., text representation, Q matrix and code representation. In each topic, different parts of the text are associated with different source codes and knowledge points. Therefore, in this paper, an attention-based LSTM architecture is designed to learn the representation of each topic as a neural network, where two attention strategies are used in this paper, one for text-source code (TCA) and one for text-knowledge points (TQA), which captures the text-source code (TCA) and text-knowledge point (TQA) associations, respectively.

In this paper, we use LSTM based architecture to learn the representation of word sequence topics of any length. The input to the LSTM network is all the data for each topic of the sequence x=(x1,x2,⋯xN) . The hidden state h of the input step of step t, is updated as: (15)it=σ(WXixt+Whiht−1+bi)ft=σ(WXfxt+Whfht−1+bf)ot=σ(WXoxt+Whoht−1+bo)ct=ftct−1tanh(WXcxt+Whcht−1+bc)ht=ottanh(ct)

Where i_·, f_·, c_·, and o_· are the input gates, forget gates, memory cells, and output gates of the LSTM respectively. w_· and b_· are the learned weight matrix and bias values.

Obviously, in each input step, x_t is a multimodal vector that combines text, knowledge points and source code, i.e: (16)xt=Wt⊕u^t⊕q^t

where ⊕ is the operation of joining two vectors into one long vector, W_t is the representation of the step t word vector in Text ET, and u^t and q^t are the representations of the source code keywords and knowledge matrices of the topic, learned by TCA and TQA, respectively.

TCA aims to capture the association between text and source code. In TCA, step t input step, the association between each keyword of text W_t and source code is first measured. Since the keywords of the source code are related to the phrases in the text section, and h_t−1 has information about the words before the input step t, it is necessary to take h_t−1 into account when measuring the associations. Then, the related source code keyword representation u^t is modeled as a vector, and the result of the weighted summation by u is denoted as: (17)u^t=∑j−1Lαjujαj=φ(uj,wt,ht−1)∑i−1Lφ(ui,wt,ht−1)φ(uj,wt,ht−1)=Vactanh(Wac[uj⊕wt⊕ht−1])

where V_ac and W_ac are the learning parameters of TCA, while in topic E, (uj,wt,ht−1) measures the association between the source code keywords u_j and w_t in step j, and α_j represents the attention parameter normalized to φ(uj,wl,hi−1) .

The goal of TQA is to capture text-knowledge point associations. Similar to TCA, the wavelet transform related knowledge point representation q in TQA can be modeled in the form of Eq. (17), where one can simply use q_j and the learned TQA parameters v_aq and W_aq instead of u_j, V_ac, and W_ac.

By attention based LSTM, the hidden state sequence h=(h1,h2,⋯hN) can be obtained in combination with the input sequence x. In addition, this approach comes from the LSTM application in natural language processing, and finally the hidden state h_N possesses the semantic information of the topic E as a whole of the input sequence x, so the semantic representation of h_N is used for r^(E), i.e., r^(E) = h_N. In addition, step t hides the state h_t to save only the information of the sequence (x1,x2,⋯xt) . Therefore, further h(E)=h is denoted as the representation of the topic E heterogeneous data, thus obtaining the unified semantic representation (r(E),h(E)) of E. 2)

Attention Mechanism (SA)

In this paper, similar attention mechanisms are designed to measure the similar parts of two exercises and learn their attentional representations. In terms of the attention mechanism, an attention matrix A is used to measure the similar parts of the input topic pair (Ea,Eb) by calculating the cosine similarity between each part of E_a and E_b and h(Ea) and h(Eb) . A∈ℝNEa×NEb can be represented as: (18)Aij=cos(hi(Ea),hj(Eb))

where 1<i<NEa , 1<j<NEb , NEa and NEb are the lengths of the word sequences in E_a and E_b respectively. hi(Ea) is the step i representation in h(Ea) and hj(Eb) is the step j representation in h(Eb) .

With Attention Matrix A, it can be found that Total Score si(Ea)=∑k=1NEbAi,k actually measures the summed similarity between the step i representation in h(Ea) and each step representation in h(Eb) . Similarly, the total score value sj(Eb)=∑k=1NEaAj,k measures the summed similarity between the step j representation in h(Eb) and each step representation in h(E0) . Therefore, the two similarity score vectors s(Ea) and s(Eb) are denoted as similarity attention representations of E_a and E_b, respectively.

For topics E_a and E_b, the semantic attentional representations of exercise topics E_a and E_b, i.e., moments h(Ea) and h(Eb) of t, can be modeled with the weighted sums of hatt(Ea) and hatt(Eb) , respectively: (19)hatt(Ea)=∑i=1NEaAi,NEahi(Ea)hatt(Eb)=∑i=1NEbANEb,jhi(Eb)

With the help of the attention mechanism, the attention matrix A can be obtained and the similar attention representations (s(Ea)) and (s(Eb)) and semantic attention representations hatt(Ea) and hatt(Eb) of the input topic pair (Ea*Eb) can be learned. 3)

Prediction Score (SL)

The goal of the prediction scoring layer is to compute the similarity score for each topic pair and find the similarity score for each pair by using the similarity score to rank the set of similar topics. By utilizing the similarity scores of it topic pairs (Ea,Eb) . First they are connected to a vector i.e. z˜ab=r(Ea)⊕r(Eb)⊕S(Ea)⊕S(Eb)⊕hatt(Ea)⊕hatt(Eb) . Then the similarity score S(Ea,Eb) is obtained by using two fully connected networks with nonlinear activation functions ReLU(x) = max(0, x) in the first one and Sigmoid function in the second one. i.e.: (20)o˜ab=ReLU(W1z˜ab+b1)S(Ea,Eb)=σ(W2o˜ab+b2)

where W₁, b₁, W₂, and b₂ are network parameters.

In this section, a pairwise loss function is specified in order to train the HANN. In the training phase, for topic E, sim(E) is used to denote the labeled similar topics and the unlabeled topics are considered as dissimilar DS(E). Considering that the similarity score of similar pairs (E,Ex) is higher than that of dissimilar pairs (E,Eds) , where E_s ∈ Sim(E), E_ds ∈ DS(E), the pairwise loss function is further constructed as: (21)L(Θ)=∑E,Es,Edmax(0,μ−(S(E,Es)−S(E,Eds)))+λΘ║Θ||2

where s(⋯,⋯) is calculated by Eq. (20). Θ is all the parameters of the HANN, and λ_Θ is a regularized hyperparameter. μ is an edge parameter that forces S(E,Es) to be greater than S(E,Eds) . In this way, HANN can be learned by directly minimizing the loss function L(Θ) using Adam [28].

In this section, we experimentally validate the effectiveness of the knowledge tracking approach based on the DKVMN-F model by comparing the performance of HRGKT with five benchmark models on four real English online education datasets, namely, ASSISTments 2009, Statics2011, ASSISTments2012, and Synthetic-5, and at the end of the ablation experiment to observe the effectiveness of the core module of the DKVMN-F model.

The benchmark models used are shown below: 1)

DKT: Modeling students’ knowledge states using RNN.

In the experiment, AUC (area under the curve) evaluation index was used to assess the performance of the model. The ROC curve, which is fully known as receiver operating characteristic curve, is plotted with the false positive rate (FPR) as the horizontal coordinate and the true positive rate (TPR) as the vertical coordinate. As the ROC curve converges more to the upper left corner, it means that the performance of the classification model performs better. The value of AUC is equal to the area under the ROC curve, and the larger the value of AUC, the better the performance of the model.

The model in this paper is experimented on a host computer with 32 GB of memory, Intel Core i7, and RTX 3060. The model randomly divides the dataset in a ratio of 7:2:1 into 3 parts: training set, validation set and test set.

The results of AUC comparison between the DKVMN-F model of this paper and the five benchmark models on the four publicly available datasets are shown in Table 1. The AUC values of the DKVMN-F model on the four datasets are 0.8669, 0.8670, 0.7757, and 0.8701, which are all better than the other five models. Relative to the other models, the DKVMN-F model achieves an average AUC value improvement of 7.86%, 5.17%, 4.49%, and 6.86% on the 4 datasets, respectively, which proves the superiority of this paper’s model.

The DKVMN-F model proposed in this paper introduces features such as the number of past answer times and repetition intervals on the basis of the DKVMN model, i.e., it integrates learning and forgetting behaviors, which are named as learning module and forgetting module, respectively. In this section, ablation experiments will be conducted to validate the performance of the learning module and forgetting module in DKVMN-F. The values of AUC for different variants of HRGKT on the four datasets are shown in Table 2. DKVMN-F (basic) denotes the baseline model without the addition of the learning module and the forgetting module, DKVMN-F denotes the final knowledge-tracking model, and the values of AUC for different variants of HRGKT are shown in Table 2. DKVMN-F (without forget) and DKVMN-F (without learning) models, while DKVMN-F (without learning) denotes the model with only learning module or forgetting module added, respectively.

As can be seen from Table 2, DKVMN-F (basic) has the lowest AUC, which is due to the fact that this model does not use learning and forgetting modules, and does not take into account past learning behaviors as well as the students’ forgetting behaviors. DKVMN-F (without forget) ignores the students’ forgetting behaviors, and its predictive performance decreases compared to that of DKVMN-F, which suggests that the factors affecting the forgetting are effective in accurately tracking students’ knowledge status, and forgetting behavior has an impact on students’ knowledge status. DKVMN-F (without learning) ignores students’ past answer information, and the predictive performance of the model is not as good as that of the DKVMN-F model, and decreases, the results indicate that the past answer information captured by the learning module is effective in improving the accuracy of the model.

In order to verify the effectiveness of the forgetting module, this paper compares the prediction of knowledge state by DKT+forget, DKVMN and DKVMN-F models, and the results of the comparison of knowledge state by different models are shown in Figure 3. Where the arrows indicate the chronological order. Ten questions answered by a randomly selected student from ASSISTments2009, which cover five knowledge points. From Fig. 3, it can be seen that DKT+forget at the moment of t=10 predicts lower results for the knowledge point k₃. Compared with DKT+forget, the probability of forgetting k₃ is lower than that of DKT+forget because DKVMN-F is able to capture the fact that Q6 at moment t=10 and Q3 at moment t=9 have the same knowledge point. at moments t=13, t=15 DKT+forget predicts lower probability of k₄, which is due to the fact that Q4 and Q10 are appearing for the 1st time in the answer sequence in the sequence of answers, and it is not possible to capture the relationship between the two questions that contain the same knowledge points. In addition, DKVMN-F has a lower prediction probability than DKVMN at moment t=16 after learning k₅ at moment t=12 without learning k₅ in between because DKVMN does not consider the forgetting factor. Therefore, it can be proved that the forgetting module is very effective in knowledge tracking.

Figure 3.

Comparison results of knowledge states of different models

In order to verify the effectiveness of the learning module in knowledge tracking, 120 knowledge points in the ASSISTments2009 dataset are clustered, using the same icon to indicate similar knowledge points, and the knowledge points are renumbered for ease of representation, and the results of the dataset clustering are shown in Figure 4.

Figure 4.

ASSISTments2009 Data clustering class

In Figure 4 the knowledge points are divided into 9 categories, which shows that the learning module is able to categorize the knowledge points of the topic well, and is able to obtain better information about the students’ previous answers and the relationship between the two, which can prove the effectiveness of the module in knowledge tracking.

1)

Data sets

In this paper, two types of datasets are designed to validate the effectiveness of sequence recommendation models. The first type is two regular sequence recommendation model datasets: MovieLens and Gowalla. MovieLens is a movie rating dataset created by the GroupLens project team. MovieLens-1M (ML-IM), which contains 1M ratings, is selected for the experiment. In addition, Gowalla is a social location-based website that allows users to share their location to realize the punch card function, and the check-in dataset in Gowalla is used in this experiment. The second category is the educational dataset ASSISTments2009.

The results of the performance comparison between the HANN model proposed in this paper and the eight baseline models listed, on the MovieLens, Gowalla, and MovieLens as well as ASSISTments2009 datasets are shown in Tables 3, 4, and 5, respectively.

In the MovieLens and Gowalla datasets, the sequential model (e.g., Caser) was found to outperform the non-sequential model (e.g., BPR) in terms of performance across the baseline methods. This observation validates the significant improvement of the sequential model for model prediction. This may be due to the fact that the sequential model better takes into account the sequential relationship between items. When comparing the different forms of sequence modeling, it is noted that FPMC models through first-order Markov chains, while CosRee achieves superior results by capturing higher-order sequence relationships. This demonstrates how considering higher order sequence relationships may help improve the accuracy of model predictions. Three different convolutional neural methods were compared in the dataset; Caser, CosRec and HANN. Caser utilizes a one-dimensional convolutional kernel to capture the evolution of a user’s interests, CosRec models this by pairwise encoding and using a two-dimensional convolution to capture more complex interactions between items. And unlike these two, the model HANN proposed in this paper introduces a contrast learning framework for heterogeneous and structured data, utilizing an LSTM based on the attention mechanism as well as pairwise loss functions to accomplish the contrast framework. This makes HANN largely outperform the previous two in terms of modeling effectiveness, and HANN achieves a performance improvement of 0.5%-7.5% on both MovieLens-1M and Gowalla datasets. This may be due to the fact that contrast learning provides a deeper representation of information when learning complex feature relationships.

The experimental results on the ASSISTments2009 dataset show that HANN outperforms other models on several evaluation metrics. Specifically, on the MAP metric, there is a gradual improvement from 0.0303 for Caser-DR to 0.0431 for HANN. On Precision@1, HANN reaches a maximum of 0.0661, which is a significant improvement compared to 0.0591 for the CosRec base model. A similar increasing trend is observed on Precision@5 and Precison@10 metrics. In terms of recall, HANN also shows the best performance, especially on Recall@5 and Recall@10, reaching 0.0218 and 0.0389, respectively. These experimental results fully demonstrate the effectiveness of adding an exercise similarity attention mechanism to the HANN model, especially the significant improvement in precision and recall, which suggests that the model has a strong potential to deal with education recommendation domain problems, demonstrating the strong potential and practical value of the model to be applied to the task of intelligent recommendation of exercises in English teaching.

This experiment is used to compare the impact of learning loss weights on the model effect, and analyze the performance change of HANN model by setting different comparison learning weights λ. The experimental results of selecting Precision@5 and Recall@5 as the comparison indexes on the ASSISTments2009 dataset are shown in Fig. 5.

Figure 5.

Influence of different comparative loss weights on Precision@5 and Recall@5

From the experimental results in Fig. 5, it can be seen that the model performance is more sensitive to specific weight coefficients. When the loss weight λ is adjusted to a lower value, especially around 0.1, the model shows better performance than other weight settings in both precision (Precision@5) and recall (Recall@5). This phenomenon suggests that lower loss weights help the model to focus more effectively on its primary task of learning, while when the weights are set to 0, the model will not use the contrast learning framework for data expansion and will be the baseline model for this paper.

The learner portrait, a tool that comprehensively describes students’ learning characteristics, interests, and cognitive styles, brings unprecedented precision to the design of English language instruction. In order to build an effective learner profile to support game task-driven instruction, we first collect multi-dimensional information on students’ historical learning data, course performance, and practical use of English.

Through data analysis and mining, students’ learning preferences, strengths and weaknesses are identified, and an accurate learner profile is drawn for each student. In the game task-driven teaching mode, this portrait information is utilized to design tasks for students that match their interests and ability levels.

Taking the 21st grade students majoring in Artificial Intelligence in the School of Computer and Software of Dalian Neusoft School of Information as an example, a game task-driven teaching design was carried out based on the learner portrait. The specific implementation steps are as follows. 1)

Analyze the learner portrait and determine the teaching objectives. Analyze the learner portraits of the participating students, including information about their learning styles, interest preferences, English proficiency, oral communication skills, and so on. Through data analysis, identify students’ commonalities and differences in order to provide them with personalized learning support and challenges. Meanwhile, the instructional objectives were designed to: (1)

Enhance students’ practical application of English.

After the completion of the English teaching tasks, a diversified evaluation method was used to assess the effectiveness of the teaching. Students’ course performance and their actual English proficiency were subjectively and objectively rated through preset grading criteria.

Taking the first-year students of non-English majors in College H as an example, comparing the English scores of the previous students, the average academic performance of the students has improved significantly. Compared with the pre-implementation period, the average grade increased by 13.6%. Among them, the number of good and above (80-100 points) increased by 8.8% and the failure rate decreased by 4.5%. In terms of classroom participation, according to teachers’ observation records, the number of students’ independent presentations increased by 22.8% after the implementation of the teaching model. At the same time, the frequency of students’ questions and discussions after class increased significantly.

Through the student satisfaction survey, it was found that 89.7% of the students were positive about the design of the teaching model, and 42.9% of them were very satisfied. Compared with the pre-implementation period, the satisfaction level increased by 8.9 percentage points. Students generally agreed that the model could better meet their English learning needs and improve learning outcomes.