An artificial intelligence recognition model for English online teaching

Abstract

The traditional English online teaching model is limited by the teaching location and the difficulty of online teaching, which prevents teachers from controlling students. In order to improve the ability of the English online teaching model to supervise and recognize the status of students, this paper proposes an English online teaching model based on artificial intelligence technology, and adopts a positioning method based on an improved deep belief network for real-time position control and status recognition for students in online learning. Moreover, this study combines intelligent algorithms to build the model structure and verify the performance of the model. The results show that the performance of the model is good. In addition, on the basis of performance testing, the recognition effect of the artificial intelligence-based student online learning recognition model constructed in this paper is recognized. The results show that the model proposed in this paper has a certain effect and meets the actual needs of intelligent teaching.

Keywords

Artificial intelligence English online teaching of artificial intelligence teaching model

1 Introduction

The role of English in the world is self-evident, and for talents in the new era, English is an indispensable skill. In order to improve the efficiency of English teaching and facilitate the timely learning of learners, the online English learning method has become an important channel for learning English.

At present, more attention must be paid to the management of artificial intelligence online teaching management systems in universities. Overall, it is closely related to the development of many fields [1]. Because of this, from a certain perspective, the development and innovation process of high-quality talent faces enormous pressure and challenges. Incorporating teaching into practice is already an important point that cannot be ignored in teaching and is also one of the development directions. The core goal of teaching is to cultivate and deepen the innovative ideas of college students and comprehensively enhance the quality and ability of students.

With the development of information technology and the widespread popularity of computer networks, online teaching of artificial intelligence has become the most popular form of distance teaching. In 2012, Harvard University and Massachusetts Institute of Technology established the EDX artificial intelligence online teaching platform. The Massive Online Open Courses (MOOC) provided by it [2] have accumulated a total of 4.5 million participants in just 4 years. Soon after, Tsinghua University, Peking University, University of California Berkeley and other major universities at home and abroad joined the EDX teaching platform. The purpose of the platform is to establish a shared education platform that combines the world’s top universities, so that learners from all over the world have the opportunity to enter the classrooms of famous schools. Through the Internet, students only need a terminal device (personal computer, tablet, mobile phone, etc.) to enjoy the courses of world-renowned schools without leaving home. The rise of the “Programming for All” movement has made more and more people realize the importance of programming. The United States has included programming courses in the K12 education system. Although MOOC teaching is very hot, it lacks personalization in programming teaching and cannot meet the needs of learners. In this context, more and more research institutions are working on the development of online programming teaching platforms. Online programming teaching platforms such as Scratch developed by MIT, Alice developed by Carnegie Mellon University, BlueJ jointly developed by University of Kent and Deakin University, and POP developed by Peking University make the form of programming learning more abundant and interesting, and they are welcomed by learners. The convenient model-driven software development method is a formal MDD development method proposed by the instructor Professor XueJinyun. The core theoretical technology of this method is the PAR method and PAR platform. It is the research result led by the research team led by Professor XueJinyun and funded by a number of national projects. The convenient model-driven software development method can effectively improve the efficiency and reliability of software development and is especially suitable for the development of complex algorithm programs and database application programs. At present, the research results have been applied in Safety-Critical software [3] in aviation and military fields. However, there are still few people mastering this innovative research result at present. The main reason is that the related theories based on formal model-driven software development are relatively complex, and the corresponding auxiliary teaching system is lacking, which limits the popularization of this method. Programming teaching and software development teaching have many things in common [4]. Therefore, combining the advantages of MOOC teaching, online programming teaching and other teaching platforms, this paper designs and implements the corresponding artificial intelligence online teaching system for the learning characteristics of convenient model-driven software development methods. The functions of the teaching system include course teaching, online modeling, online program compilation and operation, etc. The combination of these functions and course teaching helps students understand the software development method step by step. At the same time, it also facilitates the promotion of research results related to convenient model-driven software development methods.

2 Related work

Foreign teaching resource management has a long history, so the relevant research methods and concepts are worth learning and learning from. In particular, in terms of artificial intelligence and various information collection, the overall technical strength of foreign countries is very strong, and there is still a clear gap in China. The literature [5] believed that global informatization is gradually maturing. Since the birth and rise of the Internet, on the one hand, it can save massive amounts of data, and on the other hand, it can accelerate the speed of communication. In this case, teacher-student communication will be faster and more convenient. The artificial intelligence online teaching management system of colleges and universities has received a lot of support and trust, and its development is getting faster and faster [6]. Compared with the current stage, there are many limitations in the early stage of system development. The system cannot be detailed in information query, and the accuracy is low, and it cannot solve the actual problems of teaching staff [7]. As a result, it increases the daily workload and brings many negative effects. From other perspectives, the artificial intelligence online teaching management system implements a reasonable allocation of teacher authority. Teachers can perform various operations and information processing according to personal teaching ideas, which further simplifies the process and significantly reduces the difficulty [8]. In addition, as the core part of the system, background management also needs to complete the scientific design, so as to demonstrate the powerful functions [9]. At present, many teaching resource management systems still continue the traditional structure, various new demands are not paid attention to, the development of functional modules is insufficient, and the practicality is greatly reduced [10]. In fact, most system page designs are not significantly different, and overall changes are relatively small. However, the teaching information of the system can still be updated in real time. In this way, it can ensure the value of database resources and provide adequate protection for teachers and students. Under this circumstance, the teaching pressure is significantly reduced, and teaching work obtains more help [11]. The technical level of the foreign background management module is also better than the domestic level, which helps teachers to process teaching information materials and enhance work efficiency. Moreover, teachers can directly modify and update the relevant information in the database according to their needs. In addition, the background running module has a complete system and process, which can fully demonstrate its application value. In view of the current situation, the more mature system is the ATA teaching management system, which has a very high reputation abroad. As a product launched in the United States, it is also one of the common tools for computer examinations in the United States. The function of this system is very perfect. It can not only accurately simulate the teaching scene, perfectly match the real situation, but also reflect the characteristics of intelligence and informatization, and it can also choose courses by itself and answer questions/tests online. In this way, students can choose the corresponding questions to complete the test under the same difficulty. For example, it can give candidates in a certain examination room to select test questions in various subjects to make the test content more abundant and detailed [12]. However, its system maintenance requires the special personnel of the development company to deal with it. After the system is put into use, the company’s technical support is still inseparable, so there are certain limitations in practical application. In the system operation stage, the interface language is English, and there are a lot of professional nouns, so it is difficult to understand. From the perspective of students, the operation process faces many limitations and obstacles [13].

The rise of online teaching of artificial intelligence has set off a wave of online course learning all over the world. The biggest advantage of online teaching of artificial intelligence is the accumulation of a large number of high-quality courses. The teaching method of courses is mainly based on video [14]. Usually a complete video course is cut into multiple short fragments of two or three minutes, so that learners can make full use of fragmented time to learn. Although the courses on the artificial intelligence online teaching platform cover most professional courses in the higher education stage, the teaching effect of some practical courses (such as algorithmic programming) is not ideal. The main reason is that the learning of such courses requires a close combination of theory and practice, while the video-based teaching method of the artificial intelligence online teaching platform is difficult to take into account the cultivation of students’ practical skills. In addition, because the group targeted by the artificial intelligence online teaching platform is mainly students from colleges and universities, there is a lack of personalized teaching to students. Because of this, many well-known research institutions in the United States have already begun to develop online programming teaching software [15].

3 Artificial intelligence online student positioning method based on improved deep belief network

This paper proposes a positioning method based on an improved deep belief network, which is divided into two stages. First, the method collects a sample data set in the offline stage and uses the data set to construct and train a position classifier based on logistic regression. Moreover, each sample contains: reference point position coordinates, RSSI vector and sub-region number. At the same time, this method trains the corresponding deep confidence network for each sub-region. Then, in the actual positioning stage, the method uses real-time collected RSSI information to determine the sub-region to which it belongs through a position classifier, and then calls the deep confidence network corresponding to the sub-region to complete further position coordinate preprocessing [16]. Network based bonded cognitive channel are analyzed using IoT [17]. Different algorithm models were proposed in cloud based network for associating and evaluating processes [18]. Convolutional neural network based analysis were utilised for finding flood detection system [19]. Information granulation based model was optimized for various social network [20]. Virtual machine learning algorithms were utilised for provision of data intensive services [21]. Novel algorithm based approach was utilised for segmentation [22].

According to the scene of indoor positioning area in this paper, the area can be divided into two sub-areas inside and outside the room, the corridor area is marked as area A, and the area inside the room is marked as area B. Areas A and B have different scene characteristics: Area A is empty but there are mobile personnel, and Area B is equipped with tables, chairs, laboratory equipment, etc., but the movement of personnel is relatively small. If the two positioning processes are distinguished according to the characteristics of the two types of positions, and the position classifier is used as the first part of the secondary positioning method based on the improved deep confidence network and the position coordinate prediction based on the deep confidence network is more regionally targeted, it will make the location coordinate prediction based on the deep confidence network more regionally targeted, and can also obtain more accurate positioning results.

According to the binary logistic regression model, we assume that there are m APs and n terminals to be tested in the area. The terminal uses the RSSI sample set RSS ={ RSS₁, RSS₂, ⋯ , RSS_n } collected offline to train the logistic regression classifier. We assume that the RSSI sample of each terminal is x ={ x₁, x₂, ⋯ , x_m } and the classifier parameter is θ ={ θ₁, θ₂, ⋯ , θ_m }. Then, the binary logistic regression model is: $h_{θ} (x) = \frac{1}{1 + e^{- θ x}}$ (1)

Then, the binary output y of the classifier is expressed as: $y = {\begin{matrix} 0, & h_{θ} (x) < 0.5 \\ 1, & h_{θ} (x) > 0.5 \end{matrix}$ (2)

According to the above formula, the probability distribution function expression of y is: $P (y | x, θ) = h_{θ} {(x)}^{y} {(1 - h_{θ} (x))}^{1 - y}, y = 0, 1$ (3)

The log-likelihood function is inverted as the loss function of the classifier, and the parameter θ ={ θ₁, θ₂, ⋯ , θ_m } is solved by minimizing the loss function. Then, the expression of the loss function is [17]: $\begin{matrix} J (θ) = - ln L (θ) = \\ - \sum_{i = 1}^{m} [y_{i} log (h_{θ} (x_{i})) + (1 - y_{i}) log (1 - h_{θ} (x_{i}))] \end{matrix}$ (4)

Using the matrix method, it is expressed as: $J (θ) = - Y^{T} \cdot log h_{θ} (x) - {(E - Y)}^{T} \cdot log (E - h_{θ} (x))$ (5)

In the formula, E is an all 1 vector, · is an inner product, and Y is a vector composed of output y. The formula for iteratively solving the parameter θ in the loss function using the gradient descent method is expressed as follows. Among them, α is the learning rate of the gradient descent method. $θ = θ - α X^{T} (h_{θ} (x) - Y)$ (6)

The logistic regression binary classifier is constructed by the above method, and the position corresponding to the sub-region number collected offline is used to train it to obtain a position classification method based on logistic regression. The classifier can be regarded as a position classifier as a preprocessing device for improving the deep confidence network. In the actual positioning stage, the classifier is first used to determine the sub-region based on the RSSI of the target to be tested, and then the deep confidence network corresponding to the sub-region is called to perform the actual positioning process. The position classification method based on logistic regression can not only take the second classification of the position, but also can be extended to realize the multi-classification function. The specific method will not be repeated in this paper.

Deep Belief Network (DBN) is a deep learning model composed of multiple Restricted Boltzmann Machines (RBM) and a layer of BP network. The deep confidence network assigns a good initial weight to the entire network by adopting a layer-by-layer training method, so that the network can obtain the optimal network parameters only after reverse fine-tuning. The most important role in layer-by-layer training is the restricted Boltzmann machine. Before introducing the deep-confidence network, the principle of RBM is first introduced, and then the application method of the deep-confidence network is introduced in conjunction with the positioning algorithm [23].

The restricted Boltzmann machine is a typical neural network model, consisting of a visible layer v and a hidden layer h. As shown in Fig. 1, the neurons in the visible layer v and hidden layer h are connected to each other, but there is no connection between neurons in the same layer. Because RBM can obtain the higher-order correlation of neurons in the visible layer through the hidden layer, feature extraction can be performed through RBM. Neurons in RBM have two states: “activated” and “inactive”, generally represented by binary 1 s and 0 s.One of the main advantages of the restricted Boltzmann machine is that when a neuron state in the visible layer is given, the state of a neuron in the hidden layer is conditionally independent of the state of other neurons in the hidden layer. Conversely, the state of neurons in each visual layer is also independent.

Fig.1

Restricted Boltzmann machine.

The restricted Boltzmann machine is an energy-based model. The energy function of the RBM jointly configured by the visible layer vector v and the hidden layer vector h is: $\begin{matrix} E (v, h; θ) = - \sum_{i} a_{i} v_{i} - \sum_{j} b_{j} h_{j} - \sum_{ij} v_{i} w_{ij} h_{j} \\ = - a^{T} v - b^{T} h - h^{T} Wv \end{matrix}$ (7)

In the formula, θ is the parameter {W, a, b} of the RBM, a is the bias vector of the visible layer, b is the bias vector of the hidden layer, and W ∈ R^i×j is the weight matrix connecting the neurons of the visible layer and the neurons of the hidden layer. The joint distribution of v and h is: $P_{θ} (v, h) = \frac{1}{Γ (θ)} exp (- E (v, h; θ))$ (8)

In the formula, Γ (θ) is the normalization factor, and Γ (θ) = ∑_v,he^-E(v,h;θ) represents the sum of all possible states of the visible layer and hidden layer vector set. According to formula (7), formula (8) can be expressed as: $P_{θ} (v, h) = \frac{1}{Γ (θ)} exp (a^{T} v + b^{T} h + h^{T} Wv)$ (9)

After the joint probability distribution P_θ (v, h) is known, the edge distribution P_θ (v) of the visible layer vector set can be obtained by summing all the states of the hidden layer vector set [19]: $P_{θ} (v) = \frac{1}{Γ (θ)} \sum_{h} e^{- E (v, h; θ)}$ (10)

Similarly, the following results can be obtained: $P_{θ} (h) = \frac{1}{Γ (θ)} \sum_{v} e^{- E (v, h; θ)}$ (11)

Due to the structural characteristics of the RBM model: full connection between layers, no connection within layers, plus the conditional probability formula, RBM has the following important properties:

When the state of the visible layer vector is given, the activation states of neurons in the hidden layers are conditionally independent. At this time, the activation probability of the jth hidden layer neuron is: $P (h_{j} = 1 | v) = sigmoid (b_{j} + \sum_{i} v_{i} W_{ji})$ (12)

Correspondingly, when the hidden layer neuron vector is given, the activation probability of the visible layer neuron is also conditionally independent:

$P (v_{i} = 1 | h) = sigmoid (a_{i} + \sum_{j} h_{j} W_{ij})$ (13)

From the above properties, the activation function of RBM is the sigmoid function. The parameter θ ={ W, a, b } of the above RBM model can be determined by the Contrastive Divergence (CD) algorithm combined with Gibbs Sampling.

(2) Deep belief network implementing RSSI positioning method

The deep-confidence network structure used in this paper to implement the RSSI positioning algorithm is shown in Fig. 2. We assume that the deep belief network contains a visual layer v, an output layer o, and n hidden layers {h₁, h₂, ⋯ , h_n }. Among them, each upper and lower layer constitute an RBM unit. Therefore, there are n RBM units {RBM₁, RBM₂, ⋯ , RBM_n }. Among them, each RBM unit has two layers, the upper layer is the RBM hidden layer, and the lower layer is the RBM visible layer. In addition, the hidden layer of the previous RBM is stacked as the visible layer of the next RBM unit, and finally a layer of BP network is added on the top as the final output layer. The above method constitutes a deep confidence network structure for implementing RSSI positioning method [24].

Fig.2

Schematic diagram of the deep-confidence network structure implementing the RSSI positioning method.

According to the characteristics of the positioning area in this article, the complete area is divided into two sub-areas, namely the room area and the corridor area. According to the algorithm idea, it can be obtained that two deep confidence networks need to be constructed in the offline stage, and the samples collected offline are divided into two categories according to the sub-region number, and the room deep confidence network and the corridor deep confidence network are trained respectively.

Since the deep belief network is formed by stacking multiple RBM units, the output of the previous RBM unit is used as the input of the next RBM, and the highest layer of the network is the BP network layer. The offline training phase can be subdivided into two processes, namely forward unsupervised training and reverse supervised tuning training. The forward unsupervised training process adopts the unsupervised greedy algorithm from low to high layer-by-layer training to ensure that the feature vectors of the input RSSI data can be kept as much as possible when the feature information of the actual RSSI vectors is mapped to the feature space of the position coordinates. The reverse supervised tuning training means that the last layer of BP network uses supervised training method, and it carries out error back propagation from top to bottom, and then fine-tune the network model parameters globally [25 –27].

(1) Through the analysis of forward unsupervised training, we know that RBM training is actually to find a probability distribution that can best produce training samples. In other words, a probability distribution is obtained, and the probability of training samples is the largest in this distribution. Since the decisive factor for this distribution is the network parameter {W, a, b}, the training goal is to find the best network parameters. The above is the learning algorithm called Contrastive Divergence (CD) proposed by Hinton in 2002.In the CD-k algorithm (k represents the number of samples), when k = 1, only one step of Gibbs sampling can achieve a good fitting effect. Therefore, the form of CD-1 algorithm is generally adopted to fit the values of various network parameters.

Specifically, the deep confidence network is formed by stacking n layers of RBM units, so the forward unsupervised training of the deep confidence network is completed by the RBM layer-by-layer training. Among them, the visible layer of each RBM unit accepts the input signal, which may be the RSSI data of the target input under the lowest layer input, or the output signal of the RBM of the previous layer. The function of the hidden layer of each RBM unit is feature extraction. All in all, the principle of RBM is to automatically extract the best feature vector of the research problem through unsupervised learning.

(2) Reverse supervised tuning training

At the top of the deep belief network, a BP network is set to receive the output vector of the stacked RBM to implement reverse supervised training. Since the RBM unit of each layer can only ensure that the weights in its own unit are optimal for the feature vector mapping of this layer and cannot ensure that the feature vector mapping of the entire network is optimal, it is necessary to set up a BP back propagation layer to supervise the error information from top to bottom layer by layer to each RBM unit, and fine-tune the entire network. After completing the bottom-up forward unsupervised training, the stacked RBM extracts the feature vector of the RSSI data of the target to be tested, and the network parameters at this time are used as the initial connection weight matrix W and bias vector b. In order to implement the positioning algorithm using the deep-confidence network, the mapping relationship between the RSSI feature vector and the position coordinate needs to be established. To achieve this goal, a BP network layer is added on top of the stacked RBM and BP back propagation algorithm is used to perform reverse tuning training on the entire deep confidence network to establish the mapping relationship between the RSSI of the target under test and its position coordinates.

The training phase of the stacked RBM network can be regarded as the parameter initialization process of the BP network training. The combination of forward and reverse training makes the deep confidence network parameter training overcome the difficulties in the BP network training: Due to the random initialization of the weight parameters, it is easy to fall into the shortcomings of local optimization and too long training time.

BP back propagation algorithm usually chooses the mean square error (MSE) as the loss function: $J (W, b, x, y) = \frac{1}{2} {∥ a^{L} - y ∥}_{2}^{2} = \frac{1}{2} {∥ σ (W^{L} a^{L - 1} + b^{L}) - y ∥}_{2}^{2}$ (14)

Among them, a^L is the actual output vector of the network (total L layers), and y is the expected output vector of the network. The algorithm solves the connection weight matrix W and bias vector b of each hidden layer and output layer by minimizing the network output loss function J (W, b, x, y), and then completes the parameter tuning training of the entire network.

The secondary positioning method based on the improved deep confidence network completes two tasks in the training phase: one is the training of the position classifier, and the second is to construct an independent deep confidence network for each sub-region, and use the classified data samples to complete the corresponding network training. This paper obtains two deep confidence networks for the room area and corridor area. In the actual positioning stage, as a preprocessing device for improving the deep confidence network, the position classifier first collects real-time RSSI information of the target to be tested and confirms the called deep confidence network number through the classifier. After that, it inputs the RSSI information into the corresponding deep confidence network to obtain the network output result, that is, the rough position coordinates.

4 Positioning method for student dead reckoning based on artificial intelligence

Under the premise of knowing the position of the previous moment, the traditional positioning method of dead reckoning uses the motion offset angle and step length measured by the device IMU to calculate the current position. However, since the step size of each user is different, if a uniform step size is used for calculation, a larger error will inevitably occur. Therefore, this paper designed to use Recurrent Neural Networks (RNN) to predict the current location based on the user’s past location information. Through analysis, it can be seen that the trajectory information of the user in the past period contains the user’s step information. Therefore, with the help of recurrent neural network, the characteristics of the previous time series “memory” can be preserved, and the method of student dead reckoning can be improved.

The left part of Fig. 3 is the principle model of recurrent neural network. After the model is developed in time series, the right part of Fig. 3 can be obtained. The figure describes the recurrent neural network model around time slot t. The variables that appear in the figure include:

Fig.3

Principle model of recurrent neural network near time t.

x_t represents the training sample input at time slot t, similarly to x_t-1 and x_t+1;

h_t represents the hidden state of the recurrent neural network at time slot t, and h_t is determined jointly by x_t and h_t-1;

O_t represents the output vector of the recurrent neural network at time slot t, and O_t is determined by the current hidden state h_t of the network;

L_t represents the loss function of the network at time slot t, and this paper chooses the squared loss function as the loss function;

y_t represents the expected output vector of training sample x_t at time slot t;

U, W and V are the network parameter matrices of the recurrent neural network shared by the entire recurrent neural network.

For any time slot t, through the principle of recurrent neural network, we can know the hidden state h_t is determined by both x_t and h_t-1, and can be expressed as: $h_{t} = α (z_{t}) = α ({Ux}_{t} + {Wh}_{t - 1} + b)$ (15)

Among them, α () is the activation function of the recurrent neural network, which is usually the tan function, and b is the linear offset. At this time, the output vector O_t is: $O_{t} = γ ({Vh}_{t} + c)$ (16)

Among them, c is the offset between the hidden layer and the output layer, and γ () is the activation function of the output layer. The loss function of the recurrent neural network is the cross entropy between the expected output vector y_t and the actual output vector O_t. According to the actual task requirements and the content of the research in this article, the characteristics of indoor position coordinates are sought. The loss function of the recurrent neural network is set as the square loss function, which is used to measure the error between the actual output vector O_t and the expected output vector y_t, that is: $L_{t} = {(y_{t} - O_{t})}^{2}$ (17)

According to the BPTT theory and the principle of recurrent neural network, the network parameter updated during back propagation is U, W, V, b, c. Due to the characteristics of the recurrent neural network: each time slot of the network has a loss function L_t, the overall loss L is expressed as: $L = \sum_{t = 1}^{T} L_{t}$ (18)

Generally, the gradient descent method is used to complete the back propagation. In this paper, an improved algorithm, the mini-batch gradient descent method is used for reverse training.

The gradient of the overall loss function $L = \sum_{t = 1}^{T} L_{t}$ to the parameter U, W, V, b, c is calculated first.

(1) Among them, the calculation of the output layer weight and the gradient of the offset V,c is: $\frac{\partial L}{\partial c} = \sum_{t = 1}^{T} \frac{\partial L_{t}}{\partial c} = \sum_{t = 1}^{T} \frac{\partial L_{t}}{\partial o_{T}} \frac{\partial o_{T}}{\partial c}$ (19) $\frac{\partial L}{\partial V} = \sum_{t = 1}^{T} \frac{\partial L_{t}}{\partial V} = \sum_{t = 1}^{T} \frac{\partial L_{t}}{\partial o_{T}} \frac{\partial o_{T}}{\partial V}$ (20)

(2) The calculation of the gradient of the hidden layer weights and offsets U, W, b is:

It can be seen from the model of the recurrent neural network that in reverse fine-tuning, the gradient loss in a certain sequence of time slots t is determined by the gradient loss corresponding to the output of the current position and the gradient loss in the sequence time slots. We define the gradient of the hidden layer at sequence index t as: $δ_{T} = \frac{\partial L}{\partial h_{T}}$ (21)

δ^t can be derived from δ^(t+1), that is: $δ_{T} = \frac{\partial L}{\partial o_{T}} \frac{\partial o_{T}}{\partial h_{T}} + \frac{\partial L}{\partial o_{T + 1}} \frac{\partial o_{T + 1}}{\partial h_{T}}$ (22)

For δ^(T), since there are no other sequence slots behind it, there are the following results: $δ_{T} = \frac{\partial L}{\partial o_{T}} \frac{\partial o_{T}}{\partial h_{T}}$ (23)

According to δ_T, the gradient calculation expression of U, W, b is obtained as follows: $\frac{\partial L}{\partial W} = \sum_{t = 1}^{T} \frac{\partial L}{\partial h_{t}} \frac{\partial h_{t}}{\partial W}$ (24) $\frac{\partial L}{\partial U} = \sum_{t = 1}^{T} \frac{\partial L}{\partial h_{t}} \frac{\partial h_{t}}{\partial U}$ (25) $\frac{\partial L}{\partial b} = \sum_{t = 1}^{T} \frac{\partial L}{\partial h_{t}} \frac{\partial h_{t}}{\partial b}$ (26)

The above loss function $L = \sum_{t = 1}^{T} L_{t}$ trains each parameter for the gradient of the parameter U, W, V, b, c, and the network parameter U, W, V, b, c after the training are finally obtained.

According to the structural principle of the recurrent neural network, it can be seen that each time slot of the network has an output, and there is only a connection relationship between the output of the current time slot and the hidden layer unit of the next time slot, and there is a node-to-node connection relationship between each hidden unit layer. This makes the recurrent neural network have the function of “memory”. The output of each time slot is not only related to the input data of the current time slot, but also related to the output of the past time slot. The characteristics of the model described above perfectly match the principles of the student dead reckoning and positioning algorithm, so this paper chose to use the recurrent neural network to improve the student dead reckoning and positioning method.

Figure 4 shows a schematic diagram of the structure of the student dead reckoning and positioning model improved by the good time series expression ability of the recurrent neural network in this paper. It can be seen from the figure that the output result of each time slot of the network is the positioning coordinates of the student’s dead reckoning, and the input data is the movement direction offset angle collected by the device IMU in real time. By combining the position coordinates of the time slot t - 1 output, the output result of the time slot t can be obtained, and so on. The position of each time slot is determined by the movement trajectory of the previous time slots and the current direction of movement.

Fig.4

Model of student dead reckoning and positioning system based on recurrent neural network.

5 Construction of English online model of artificial intelligence recognition

Based on the above artificial intelligence algorithm, this study builds an online education model that can intelligently identify students. It can locate students, determine whether students are in a prescribed learning place, and distinguish whether students are acting on behalf of others in online learning. The system model constructed in this paper is shown in Fig. 5.

Fig.5

Access control model diagram.

Figure 5 shows the access control model diagram. In the model, when intercepting a user’s request, we can use Struts and page technology to restrict the interception. At the same time, after intercepting the request information in the background, the user’s role information is extracted and transferred to the access control stage to determine the role’s use authority. There is an association relationship mapping between roles and permissions, and there is an association relationship mapping between users and roles. Under such a dual association relationship, the user’s operation authority on resources is locked. The user is the owner or subject of the authority, which is separated from the beginning and then associated together. The role-permission list is the carrier of the association of roles and permissions. In fact, the authorization process is completed, and the use permissions are specified for each role. When it is necessary to verify the user’s role permissions, it is determined by querying the role-permission list, which plays a vital role in the entire RBAC model.

From the actual situation, the relationship between the response time of the system and the request made by the user is extremely close. In addition, the system can be accessed by other methods when logging in. Internet or optical fiber can support the completion of access operations. However, in this process, it is worth noting that due to differences in network environment and different network providers, there will be differences in network speed. Network differences will also affect the operation of the system. Generally speaking, if the user only performs relatively simple operations during the login process, the response time used will not be too long. However, if the user downloads or transfers the file, the response time will be appropriately extended. When users use the system, they all hope that the system has strong operability and can respond quickly, so that the waiting time can be shortened, and the use efficiency is effectively strengthened.If the system’s response time is too slow, then users may give up because of wasting too much time. Therefore, when conducting tests, we must also pay attention to this aspect. If we find that the waiting time is too long in the detection process, we can improve it. By improving the response speed to reduce the user’s waiting time, so that users can have a better operating experience. Regarding the system login stress test, this study uses the following methods to achieve: At intervals of 2 s, 5 users are added to the system. Each user logs in normally and operates for 5 minutes, and then releases at a rate of 10 users within 1 s. The final results are shown in Table 1 and Table 2, and the response time is shown in Fig. 6.

Table 1

Statistical table of response time

Concurrent number	Response time	Fastest response	Slowest response
2000	2.23	0.22	3.60
5000	3.49	0.38	4.67
20000	4.67	0.47	7.10
25000	11.87	0.64	13.19

Table 2

Concurrent number	Clicks/S	Average throughput
2000	213.4	1125776
5000	204.7	1127009
20000	208.5	1126142
25000	206.4	1126542

Fig.6

Statistical diagram of response time.

It can be seen from the above table that during the system test, task assignments are all done with related software systems and servers (taking absolute concurrency as a multi-user processing method). According to the system response data during the concurrent task execution phase, it is not difficult to find that if the amount of concurrent operations is less than 5000, the response is extremely fast, and the entire process is very smooth.

Based on the performance test, the recognition effect of the artificial intelligence recognition student online learning model constructed in this paper is recognized. In the experiment, 60 sets of data are set, and each set of data has 100 students’ online learning actions. Through the model, the actual learning status of the students is measured, and the results are shown in Table 3 and Fig. 7.

Table 3

Statistical table of model recognition accuracy rate of students’ learning actions and states

NO.	Recognition accuracy(%)	NO.	Recognition accuracy(%)
1	98	31	93
2	97	32	93
3	99	33	94
4	95	34	100
5	96	35	94
6	97	36	97
7	100	37	99
8	93	38	96
9	91	39	93
10	97	40	100
11	98	41	96
12	97	42	95
13	93	43	96
14	93	44	92
15	99	45	92
16	99	46	95
17	100	47	92
18	98	48	94
19	92	49	95
20	93	50	96
21	93	51	93
22	98	52	97
23	97	53	95
24	91	54	100
25	99	55	94
26	94	56	97
27	93	57	100
28	98	58	93
29	97	59	95
30	96	60	96

Fig.7

Statistical diagram of model recognition accuracy rate of students’ learning actions and states.

As shown in Fig. 7, the accuracy of the model’s recognition of students’ learning actions and states is above 90%. It can be seen that the model proposed in this paper has a certain effect and meets the actual needs of intelligent teaching.

6 Conclusion

In order to improve the English online teaching model’s ability to supervise and recognize students, this paper proposes an English online teaching model based on artificial intelligence technology. The principle of the traditional positioning method is to match the database with the RSSI vector at the target to be measured. In the offline stage, clustering algorithms such as attractor propagation algorithms can be used to reduce the complexity of the matching database. In the online stage, the KNN algorithm, compressed sensing algorithm, etc. can be used to achieve database matching and obtain the final position coordinates. This paper proposes a positioning method based on improved deep belief networks. The method is divided into two stages. First, the sample data set is collected in the offline stage, and each sample contains: reference point position coordinates, RSSI vector, and sub region number. After that, the data set is used to construct and train a position classifier based on logistic regression. At the same time, the corresponding deep confidence network is trained for each sub-region. In addition, this study combines intelligent algorithms to build the model structure and verify the performance of the model. The results show that the performance of this model is good. On the basis of performance testing, the recognition effect of the artificial intelligence-based student online learning recognition model constructed in this paper is recognized. The results show that the model proposed in this paper has a certain effect and meets the actual needs of intelligent teaching.

References

Abdulsahib

G.M.

and Khalaf

O.I.

, Comparison and Evaluation of Cloud Processing Models in Cloud-Based Networks [J], International Journal of Simulation–Systems, Science & Technology 19(5) (2018).

Al-Tamimi

, Revisiting acoustic correlates of pharyngealization in Jordanian and Moroccan Arabic: Implications for formal representations [J], Laboratory Phonology 8(1) (2017), 1–40.

Choi

H.N.

, Byun

S.W.

and Lee

S.P.

, Discriminative Feature Vector Selection for Emotion Classification Based on Speech [J], Transactions of the Korean Institute of Electrical Engineers 64(9) (2015), 1363–1368.

Raj

E.D.

, Manogaran

, Srivastava

and Wu

, Information Granulation-Based Community Detection for Social Networks [J], IEEE Transactions on Computational Social Systems.

Niaz

, Khalid

, Ullah

, Aslam

, Raza

and Priyan

M.K.

, A bonded channel in cognitive wireless body area network based on IEEE 802.15.6 and internet of things [J], Computer Communications 150 (2020), 131–143.

Haderlein

, Döllinger

, Matoušek

, et al., Objective voice and speech analysis of persons with chronic hoarseness by prosodic analysis of speech samples[J], Logopedics Phoniatrics Vocology 41(3) (2015), 106–116.

Herbst

C.T.

, Hertegard

, Zangger-Borch

, et al., Freddie Mercury—acoustic analysis of speaking fundamental frequency, vibrato, and subharmonics[J], Logopedics Phoniatrics Vocology 42(1) (2016), 1–10.

Hill

A.K.

, Cárdenas

R.A.

, Wheatley

J.R.

, et al., Are there vocal cues to human developmental stability? Relationships between facial fluctuating asymmetry and voice attractiveness [J], Evolution & Human Behavior 38(2) (2017), 249–258.

Hossain

M.S.

and Muhammad

, Healthcare Big Data Voice Pathology Assessment Framework [J], IEEE Access PP(99) (2016), 1–1.

10.

Hsu

C.C.

, Cheong

K.M.

, Chi

T.S.

, et al., Robust Voice Activity Detection Algorithm Based on Feature of Frequency Modulation of Harmonics and Its DSP Implementation[J], IEICE Transactions on Information and Systems E98.D(10) (2015), 1808–1817.

11.

Kang

T.G.

and Kim

N.S.

, DNN-Based Voice Activity Detection with Multi-Task Learning [J], Ieice Transactions on Information & Systems E99.D(2) (2016), 550–553.

12.

Kumar

P.H.

and Mohanty

M.N.

, Efficient Feature Extraction for Fear State Analysis from Human Voice [J], Indian Journal of Science & Technology 9(38) (2016), 1–11.

13.

Laukka

, Elfenbein

H.A.

, Thingujam

N.S.

, et al., The Expression and Recognition of Emotions in the Voice Across Five Nations: A Lens Model Analysis Based on Acoustic Features [J], Journal of Personality & Social Psychology 111(5) (2016), 686.

14.

Leeman

, Mixdorff

, O’Reilly

, et al., Speaker-individuality in Fujisaki model f0 features: Implications for forensic voice comparison[J], International Journal of Speech Language and the Law 21(2) (2015), 343–370.

15.

Anbarasan

, Muthu

B.A.

, Sivaparthipan

C.B.

, Sundarasekar

, Kadry

, Krishnamoorthy

, Dinesh Jackson Samuel

and Antony Dasel

, Detection of flood disaster system based on IoT, big data and convolutional deep neural network [J], Computer Communications 150 (2020), 150–157.

16.

Malallah

F.L.

, Saeed

K.N.Y.M.G.

, Abdulameer

S.D.

, et al., Vision-Based Control By Hand-Directional Gestures Converting To Voice[J], International Journal of Scientific & Technology Research 7(7) (2018), 185–190.

17.

Woźniak

and Połap

, Voice recognition through the use of Gabor transform and heuristic algorithm [J], Nephron Clinical Practice 63(2) (2017), 159–164.

18.

Mohan

, Hamilton

, Grasberger

, et al., Realtime voice activity and pitch modulation for laryngectomy transducers using head and facial gestures[J], Journal of the Acoustical Society of America 137(4) (2015), 2302–2302.

19.

Sleeper

, Contact effects on voice-onset time in Patagonian Welsh [J], Acoustical Society of America Journal 140(4) (2016), 3111–3111.

20.

Mousavizadeh

, Maghooli

, Fatemizadeh

, et al., Liveness Detection in Face Identification Systems: Using Zernike Moments and Fresnel Transformation of Facial Images[J], Indian Journal of ence and Technology 8(8) (2015), 523.

21.

Duong

N.Q.K.

and Duong

H.T.

, A Review of Audio Features and Statistical Models Exploited for Voice Pattern Design [J], Computer Science 03(2) (2015), 36–39.

22.

Nidhyananthan

S.S.

, Muthugeetha

and Vallimayil

, Human Recognition using Voice Print in LabVIEW[J], International Journal of Applied Engineering Research 13(10) (2018), 8126–8130.

23.

Orlandi

, Garcia

C.A.R.

, Bandini

, et al., Application of Pattern Recognition Techniques to the Classification of Full-Term and Preterm Infant Cry[J], Journal of Voice 30(6) (2015), 656–663.

24.

Prathik

, Anuradha

and Uma

, A Novel Algorithm for Soil Image Segmentation using Colour and Region Based System [J], International Journal of Innovative Technology and Exploring Engineering 8(10) (2019), 3544–3550.

25.

Rhodes

, Aging effects on voice features used in forensic speaker comparison [J], International Journal of Speech Language & the Law 24(2) (2017), 177–199.

26.

Sarria-Paja

, Senoussaoui

and Falk

T.H.

, The effects of whispered speech on state-of-the-art voice based biometrics systems [J], Canadian Conference on Electrical and Computer Engineering 2015 (2015), 1254–1259.

27.

Wang

, Zhou

, Hsu

C.-H.

, Xiao

and Yang

, Provision of data-intensive services through energy-and qos-aware virtual machine placement in national cloud data center [J], IEEE Transactions on Emerging Topics in Computing 4(2), 290–300.