Research on innovative strategies of college students’ English teaching under the background of artificial intelligence

In the present day learning environment, the strategies of English teaching for students under artificial intelligence have stepped into a mature and innovative stage. Table 1 shows the introduction of artificial intelligence in school teaching in various countries [1–5].

There is an urgent need to increase cross-border talent training in the field of education [6]. In July 2017, the American Association for the Advancement of Science published a special issue of Science, in which was provided a detailed explanation of the practical application of artificial intelligence in education [7]. Arnaue [8] believes that artificial intelligence can be applied in the field of education. Russell and Norvig [9] point out in their research that the development of an educational method based on artificial intelligence can better assist teachers in teaching, artificial intelligence can help teachers realise personalised teaching and intelligent devices in the intelligent age can monitor and record students’ learning process, learning path and even learning choice in real time. Teachers can use an analysis engine to monitor the learning progress of each student in real time and provide personalised guidance, so as to better help students solve problems according to their pre-set methods and complete learning tasks efficiently and on time. Popenici [10] pointed out in his research that artificial intelligence can greatly improve students’ learning situation, evaluation of course quality, students’ adaptive learning and recommendation learning, with highly noticeable effects. Li Hui [11] believed that although artificial intelligence education would have a significant impact on the traditional teaching methods, the teaching profession would not disappear with it. China has used artificial intelligence equipment in teaching over the years, as shown in Figure 1.

Fig. 1

To improve China’s national image, it is necessary to export Chinese culture effectively as the context for Western understanding of China, and the key to export Chinese culture effectively is to improve the quality of English teaching for college students. It has become a social and cultural hot spot advancing with the times [12–17]. As a kind of knowledge and skill urgently needed by the society, English teaching has been attached importance to by the education department of our country, and the importance of college English teaching is increasingly prominent. In today’s world, economic, information and people-to-people exchanges among countries are more dynamic than ever. The cultivation of translation ability in college English teaching is of great significance to the whole exercise of college English teaching [18–21]. Therefore, college English teaching innovation strategies should be constantly reformed. With 1.4 billion Chinese native speakers, China has a huge market for learning English.

The research on the innovative strategies of English teaching for college students based on artificial intelligence has gradually come into view and become a hot topic for scholars at home and abroad [22–27]. This paper adopts the model algorithm combined with logistic regression and neural network to study in detail the means to make full use of these ‘high technologies’ to achieve innovative strategies in college English teaching.

Developing innovative strategies in college English teaching using artificial intelligence is one of the emerging disciplines, and while it belongs to artificial intelligence, it may also be referred to as a cross discipline. It aims mainly to have the characteristic of extensibility. Further, MoNiHua artificial intelligence technology research and application, such as the development of machines that can simulate intelligent human behaviour including language, logic, space, body movement, interpersonal relationships, etc. provides great convenience. In the English teaching of college students, there are many applications of artificial intelligence, and many teaching links and teaching evaluation cannot do without the help of artificial intelligence [28–35].

Logistic regression is one of the main methods used in students’ English teaching research innovation strategy model. Logistic regression model is based on linear regression, and through the use of logical function, it becomes possible to estimate the probability using which the relationship between the classification dependent variable and one or more independent variables can be measured. For the innovation strategy model of college English teaching, its regression probability is represented by conditional probability distribution P(Y | x), the value of random variable Y is 0 or 1, and the value of random variable X is real; accordingly, the conditional probability distribution of binomial logistic regression model is as follows: 1 P(Y=1∣x)=exp⁡(ω⋅x+b)1+exp⁡(ω⋅x+b) \begin{equation} P\left( Y = 1\mid x\right) = \frac{{\exp \left( \omega \cdot x + b\right)}}{{1 + \exp \left( \omega \cdot x + b\right)}} \end{equation} 2 P(Y=0∣x)=11+exp⁡(ω⋅x+b) \begin{equation} P\left( Y = 0\mid x\right) = \frac{1}{{1 + \exp \left( \omega \cdot x + b\right)}} \end{equation} where x ∈ Rⁿ is the input, Y ∈ {0, 1} is the output, ω ∈ Rⁿ is the weight vector and b ∈ R is the parameter. The ratio of the probability of occurrence of an event to the probability density of its non-occurrence is again expressed in terms of a probability, and this logistic regression logarithmic probability is expressed as: 3 log⁡P(Y=1∣x)1−P(Y=1∣x)=ω⋅x \begin{equation} \log \frac{{P\left( Y = 1\mid x\right)}}{{1 - P\left( Y = 1\mid x\right)}} = \omega \cdot x \end{equation}

In a regression model, the logarithm of output Y = 1 is a linear function of input x. The maximum likelihood estimation method is used to find the weight vector ω. If P(Y=1∣x)=π(x) $P\left( Y = 1\mid x\right) = \pi (x)$, P(Y=0∣x)=1−π(x) $P\left( Y = 0\mid x\right) = 1 - \pi (x)$. In order to prevent over-fitting and facilitate calculation, the logarithmic likelihood function after introducing the regular term λ∑j=1nωj2 $\lambda \sum\limits_{j = 1}^n {\omega _j^2}$ and adding the coefficient 1/n is: 4 J(ω)=−1n∑i=1N[yi(ω⋅xi)−log⁡(1+exp⁡(ω⋅xi))]+λ∑j=1nωj2 \begin{equation} J(\omega ) = - \frac{1}{n}\sum\limits_{i = 1}^{N} \left[ {{y_i}\left( {\omega \cdot {x_i}} \right) - \log \left( {1 + \exp \left( {\omega \cdot {x_i}} \right)} \right)} \right] + \lambda \sum\limits_{j = 1}^n {\omega _j^2} \end{equation}

Assuming that ωwith the smallest J(ω) is ω¯ $\bar \omega$, the learned logistic regression model is: 5 P(Y=1∣x)=exp⁡(ω¯⋅x)1+exp⁡(ω¯⋅x) \begin{equation} P\left( Y = 1\mid x\right) = \frac{{\exp (\bar \omega \cdot x)}}{{1 + \exp (\bar \omega \cdot x)}} \end{equation} 6 P(Y=0∣x)=11+exp⁡(ω¯⋅x) \begin{equation} P\left( Y = 0\mid x\right) = \frac{1}{{1 + \exp (\bar \omega \cdot x)}} \end{equation}

Two groups of variables are set in the model. One group is the state variable {y1,y2,…,yn} $\left\{ {{y_1},{y_2}, \ldots ,{y_n}} \right\}$ called implicit variable, and yi∈Y ${y_i} \in Y$ represents the system state at the ith moment. The other group is the observation variable {x1,x2,…,xn} $\left\{ {{x_1},{x_2}, \ldots ,{x_n}} \right\}$, where xi∈X ${x_i} \in X$ represents the observed value at time i. The joint probability distribution of all variables is: 7 P(x1,y2,…,xn,yn)=P(y1)P(x1∣y1)∏i=2nP(yi∣yi−1)P(xi∣yi) \begin{equation} P\left( {{x_1},{y_2}, \ldots ,{x_n},{y_n}} \right) = P\left( {{y_1}} \right) P\left( {{x_1}\mid {y_1}} \right)\prod\limits_{i = 2}^{n} P \left( {{y_i}\mid {y_{i - 1}}} \right) P\left( {{x_i}\mid {y_i}} \right) \end{equation}

Based on the above two groups of variables, given the training sample set (x_i,y_i) and the hyperplane (w⋅x)+b=0 $(w \cdot x) + b = 0$, the constraining condition yi[(w⋅xi)+b]≥1 ${{\text{y}}_{i}} \left[ \left( {w \cdot {x_i}} \right) + b\right] \ge {{1}}$ needs to be satisfied so that the classification plane has the classification interval and can correctly classify all samples. It can be calculated that the classification interval is 2‖w‖ $\frac{{\text{2}}}{{\left\| w \right\|}}$, and then the beam optimisation problem of the group is constructed and solved: 8 minϕ(w)=12‖w‖2=12(w′⋅w) \begin{equation} \min \phi (w) = \frac{1}{2}{\left\| w \right\|^2} = \frac{1}{2}\left( {{w^\prime } \cdot w} \right) \end{equation}

The Lagrange function is introduced to calculate the optimal weight w and the optimal bias b, so as to obtain the optimal classification hyperplane (w⋅x)+b=0 $(w \cdot x) + b = 0$ and the optimal classification function: 9 f(x)=sgn{(w∗⋅x)+b∗}=sgn{(∑j=1laj∗yj(xj⋅xi))+b∗},x∈Rn \begin{equation} f(x) = {\text{sgn}} \left\{ {\left( {{w^*} \cdot x} \right) + {b^*}} \right\} = {\text{sgn}} \left\{ {\left( {\sum\limits_{j = 1}^l {a_j^*} {y_j}\left( {{x_j} \cdot {x_i}} \right)} \right) + {b^*}} \right\},\quad x \in {R^n} \end{equation}

In order to improve robustness and fault-tolerance, a convolutional neural network algorithm was introduced into the model. The input of the convolutional neural network is usually the original image X, and H is used to represent the feature mapping of layer i of the convolutional neural network H₀ = X. Assuming that H_i is the convolution layer, H_i as the weight matrix can be expressed as: 10 Hi=f(Hi−1⊗Wi+bi) \begin{equation} {H_i} = f\left( {{H_{i - 1}} \otimes {W_i} + {b_i}} \right) \end{equation}

After the convolution layer is the pooling layer, also known as the down-sampling layer. Assuming H is the pooling layer, H_i can be expressed as: 11 Hi=subsampling(Hi−1) \begin{equation} {H_i} = {\mathop{\rm subsampling}\nolimits} \left( {{H_{i - 1}}} \right) \end{equation}

Essentially, convolutional neural network is a mathematical model that maps the original matrix H₀ into a new feature expression (Y) after data transformation or dimensionality reduction at multiple levels: 12 Y(i)=P(L=li∣H0;(W,b)) \begin{equation} Y(i) = P\left( {L = {l_i}\mid {H_0};(W,b)} \right) \end{equation}

In order to improve model performance, a logical sigmoid function is generally adopted: 13 f(x)=11+e−x \begin{equation} f(x) = \frac{1}{{1 + {e^{ - x}}}} \end{equation}

The hyperbolic tangent sigmoid function (Tanh) and the rectified linear unit (ReLU) serve as the excitation function of each neuron at the full connection layer: 14 f(x)=tanh⁡(x) \begin{equation} f(x) = \tanh (x) \end{equation} 15 f(x)=max(x,0) \begin{equation} f(x) = \max (x,0) \end{equation}

When differentiating the parameters of the teaching model, the input characteristics multiplied by the convolution kernel should be considered. xu+i−1,v+j−1l−1 $x_{u + i - 1,v + j - 1}^{l - 1}$ is the partial input element multiplied by Wijl $W_{ij}^l$, η is the learning rate and the weight of the regression function is updated as follows: 16 Wijl=Wijl−η∑u∑v(δuvlxu+i−1,v+j−1l−1) \begin{equation} W_{ij}^l = W_{ij}^l - \eta \sum\limits_u {\sum\limits_v {\left( {\delta _{uv}^lx_{u + i - 1,v + j - 1}^{l - 1}} \right)} } \end{equation}

Since each output value of the channel is related to the bias term, the gradient of the bias term is the sum of the sensitivity elements of the channel, and then the bias term is updated as follows: 17 bjl=bjl−η∑u,v(δjl)u,v \begin{equation} b_{j}^{l} = b_{j}^{l} - \eta \sum\limits_{u,v} {{{\left( {\delta _j^l} \right)}_{u,v}}} \end{equation}

It is assumed that the results of m sets of college students’ English teaching innovation strategies are distributed on n nodes. At the moment t, pheromone intensity on the path between node i and node j, and its update mode, are 18 τij(t+1)=(1−ρ)τij(t)+Δτij(t) \begin{equation} {\tau _{ij}}(t + 1) = (1 - \rho ){\tau _{ij}}(t) + \Delta {\tau _{ij}}(t) \end{equation} where ρ represents pheromone volatilisation coefficient; Δτij(t) $\Delta {\tau _{ij}}(t)$ represents the sum of pheromone increment on the path between node i and node j, and the specific calculation equation is as follows: 19 Δτij(t)=∑k=1mτijk(t) \begin{equation} \Delta {\tau _{ij}}(t) = \sum\limits_{k = 1}^m {\tau _{ij}^k} (t) \end{equation}

The graded probability of the results of the k set of college students’ English teaching innovation strategies is: 20 pijk(t)={[τij(t)]α[ηij(t)]β∑l∉Jk[τil(t)]α[ηil(t)]β,j∉Jk0,otherwise \begin{equation} p_{ij}^k(t) = \begin{cases} \frac{{{{\left[ {{\tau _{ij}}(t)} \right]}^\alpha }{{\left[ {{\eta _{ij}}(t)} \right]}^\beta }}}{{\sum\limits_{l \notin {J^k}} {{{\left[ {{\tau _{il}}(t)} \right]}^\alpha }} {{\left[ {{\eta _{il}}(t)} \right]}^\beta }}}, & j \notin {J^k}\\ 0, & \text{otherwise} \end{cases} \end{equation} where J^k represents the grade of the results of the k set of college students’ English teaching innovation strategies; and α and β represent adjustable parameters.

In the comprehensive Eqs (18)–(20), the output vector X_T and the hidden layer value O_t and S_t of the recurrent neural network at time t are as follows: 21 Ot=g(V⋅St) \begin{equation} {O_t} = g\left( {V \cdot {S_t}} \right) \end{equation} 22 St=f(U⋅Xt+W⋅St−1) \begin{equation} {S_t} = f\left( {U \cdot {X_t} + W \cdot {S_{t - 1}}} \right) \end{equation}

Accuracy is a commonly used evaluation index for the innovation strategy model of college English teaching. Assuming that the learned model is Y=f^(X) $Y = \hat f\left( X \right)$, the accuracy of the common test data set would be as follows: 23 Accuracy=1N′∑i=1N′I(yi=f^(xi)) \begin{equation} \text{Accuracy} = \frac{1}{{{N^\prime }}}\sum\limits_{i = 1}^{{N^\prime }} I \left( {{y_i} = \hat f\left( {{x_i}} \right)} \right) \end{equation} where N^′ is the test sample size, and I is the indicator function.

Precision indicates how many of the samples with positive predicted values are really positive samples. The definition formula is shown in Eq. (24): 24 Precision=TPTP+FP \begin{equation} \text{Precision} = \frac{{TP}}{{TP + FP}} \end{equation}

Recall is defined as Eq. (25): 25 Recall=TPTP+FN \begin{equation} \text{Recall} = \frac{{TP}}{{TP + FN}} \end{equation}

According to the traditional training depth model, the innovative strategy model of college English teaching uses the stochastic gradient descent algorithm for training parameters. According to the classification objective definition, the cross entropy loss function is selected: 26 EN=−1N∑n=1N∑k=1ctknlog⁡ykn \begin{equation} {E^N} = - \frac{1}{N}\sum\limits_{n = 1}^N {\sum\limits_{k = 1}^c {t_k^n} } \log y_k^n \end{equation} where N represents the number of samples, c represents the number of categories, tkn $t_k^n$ represents the true category of the nth sample and ykn $y_k^n$ represents the predicted result of the nth sample. The goal of the training model is to minimise the loss function. We define the output of layer Z as: 27 \begin{align} {x^l} &= f\left( {{u^l}} \right)\\ \nonumber {u^l} &= {\omega ^l}{x^{l - 1}} + {b^l} \end{align} where f represents the activation function, x^l–1 represents the output result of L – 1 layer, as well as its input for L layer, l represents the weight of L layer, and b represents the bias of L layer.

For the phenomenon of category imbalance, the more balanced evaluation index is receiver operating characteristic (ROC), which is also called subject operating characteristic. ROC does not vary with the distribution of positive and negative samples in the test set. ROC consists of two parameters: TRP and FPR. The definition equations of TRP and FPR are as follows: 28 TPR=TPTP+FN \begin{equation} TPR = \frac{{TP}}{{TP + FN}} \end{equation} 29 FPR=FPTN+FP \begin{equation} FPR = \frac{{FP}}{{TN + FP}} \end{equation}

This chapter first introduces the logistic regression model in detail and analyses the probability of college students’ English teaching innovation strategy model by using logical function to estimate the probability. In order to improve robustness and fault-tolerance, a convolutional neural network algorithm was introduced into the model. Finally, it introduces E_N, the commonly used loss function of alternative cross entropy, and the evaluation index is ROC, which mainly involves two parameters: TRP and FPR.

However, for the present, universities can supplement their teaching content with artificial intelligence. Various platforms based on artificial intelligence can not only contain the contents of textbooks, but also include many network teaching courseware and video materials. Students can not only study in class, but also learn English content through portable electronic devices such as mobile phones and tablets. Different teaching methods and a large amount of teaching content enable students to choose at will, and compared with traditional textbooks, the teaching content is more diverse. The English teaching contents and method of colleges and universities were imported into the innovative strategy model as parameters, and factors and weights affecting the quality of college English teaching were analysed and calculated, as indicated in Table 2.

As can be seen from column 1 of Table 2, the weight of teaching methods accounts for 54%. This means that the effectiveness of teaching will be affected by the way the English teacher teaches. Due to the characteristics of the English language, teaching methods should be regarded as the core of English teaching.

As can be seen from column 2 of Table 2, the weight of teaching environment in English teaching accounts for 31%. This means that the teaching effect will be affected by the classroom environment of teachers and students. Due to the characteristics of the English language, the focus of the teaching environment should be the communication between teachers and students.

As can be seen from column 3 of Table 2, the weight proportion of students’ independent learning ability is 11%. This means that the process of acquiring knowledge is learning. Knowledge should not be taught by teachers, but should be acquired by learners through a series of autonomous learning actions conducted by using necessary learning materials in certain situations and with the help of others (including learning partners and teachers, etc.). Learners’ ability to learn about the meaning of knowledge based on their own experience determines how much knowledge they acquire.

The above factor analysis involves three kinds of simulation results based on teaching methods as well as a teaching environment, both of which are in turn based on artificial intelligence. The weight of teaching quality improved by simulation analysis innovation strategy is shown in Figure 2.

Fig. 2

From Figure 2, we perceive that the weight of teaching methods based on artificial intelligence is the highest at 73%. This means that teachers can expand the scope of students’ English application according to the interactive characteristics that have been ascertained based on artificial intelligence. For example, relevant teaching software or public platforms can be used to release teaching tasks in advance. Teachers can feedback the learning progress of students through artificial intelligence devices according to their learning status, and students can use artificial intelligence devices to learn. Body language is one of the important teaching methods in classroom teaching, which can achieve a certain coherence of discourse meaning. Teachers should consider the course content before class and coordinate with the corresponding pronunciation and intonation and body movements.

Figure 2 shows that the weight of the teaching environment based on artificial intelligence is 69%. This means that teachers can create multi-modal problem situations such as games, practice, competitions and debates when they organise teaching activities. Teachers can appropriately integrate the elements of interpretation or listening classes, which will often produce unexpected positive effects on college English teaching.

Figure 2 shows that the weight of autonomous learning based on artificial intelligence is 57%. Therefore, college English teachers should give prominence to students’ dominant position in English teaching, and the teaching arrangement should always centre on students’ independent learning. This ‘student-independent learning-centred theory’ allows students to learn, think and create independently, rather than just surround themselves with teachers. Teachers need to increase the opportunities for students to speak English in class, leave the stage for students, increase the sense of participation of students and encourage them to exercise their subjective initiative.

At present, the fourth technological revolution with artificial intelligence as the core is changing people’s life, work, study and other aspects in an unprecedented situation, opening a curtain of profound impact on human social life. All kinds of social phenomena and behaviours of people can be ‘digitised’. These digitised phenomena and behaviours can be collected, stored, analysed and utilised, a collection of functions that greatly exceeds the traditional ability of information acquisition and interpretation. It provides a strong support for more accurate understanding of demand, grasping trends, providing services and predicting development. This study proves that using artificial intelligence to innovate English teaching strategies can achieve good teaching effects, primarily by way of changing students’ attitude towards learning English. There is a need to consciously guide students to pay attention to the relevant field of knowledge, so that students expand the scope of knowledge; and at the same time, an improvement in the overall quality can also be achieved, in other words a comprehensive betterment can happen encompassing not only communication skills in English but also the effectiveness and skill with which various hitherto unexplored subjects (whose body of literature might be principally in English) are studied and assimilated. The specific conclusions are as follows:

Through the college students’ English teaching innovation strategy model, the research shows that the weight of teachers’ teaching methods accounts for 54%. The weight of English teaching environment accounted for 31%. The weight of students’ independent learning ability is 11%.