Optimization Design of Artificial Intelligence Aids in Higher Vocational English Classroom Teaching under the Concept of Industry-Teaching Integration

Artificial Intelligence is based on programs, data and algorithms, which enable robots to have intelligent recognition, cognition, analysis and decision-making by simulating human intelligence [1]. At present, artificial intelligence is widely used in the field of education, and concepts such as “intelligent education” and “artificial intelligence in education” have been derived [2]. Among them, “intelligent education” is education from the perspective of informatization, and “artificial intelligence in education” uses artificial intelligence as a means of education [3]. Combined with teaching practice, artificial intelligence uses information technology as the main means to enhance education through the application of educational robots, intelligent teaching platforms, and intelligent evaluation systems, and promote the sustainable development of education [4-5].

At present, artificial intelligence technology has become an important teaching aid technology in the field of education. In English teaching activities in colleges and universities, through the use of AI technology, it can provide students with personalized, intelligent and efficient teaching services, improve students’ learning effect and experience, and at the same time, it can also reduce the burden of teachers’ teaching and promote the fair development of education [6-8]. The introduction of artificial intelligence technology provides new ideas and methods for teaching English courses in colleges and universities, and can promote the innovation of teaching mode, improve teaching quality and efficiency, and promote the personalized development of students [9-10].

Therefore, English teachers in colleges and universities should actively explore and apply artificial intelligence technology to promote the reform and development of English teaching [11]. Traditional English teaching is limited by textbooks and teaching aids, and the teaching content is relatively single, making it difficult to meet the learning needs of different students [12]. By using AI technology, it enables teachers to easily access and integrate a large number of digitized teaching materials and teaching resources [13]. Teachers can use online resource libraries to select teaching materials suitable for different teaching objectives and student levels, and even customize personalized learning content according to students’ interests and needs [14-15]. These resources may include e-books, interactive exercises, video lectures, online seminars, etc., which greatly enrich the teaching tools [16].

In traditional teaching practice, teachers need to invest a lot of time in homework correction, grade statistics, and learning situation analysis [17]. By utilizing AI technology, teachers can efficiently complete their daily teaching with the help of intelligent teaching systems, thus releasing more time and energy to focus on the design of teaching programs and the innovation of teaching methods, and finally transforming their role orientation [18-19]. Traditionally, English teachers usually play the role of knowledge transmitters, while students are often in a passive state of acceptance. However, with the strong support of AI technology, the role of teachers is gradually changing to that of a guide and partner in the students’ learning process [20-21]. Teachers are able to use advanced AI technology tools to achieve real-time interaction with students, answer students’ questions, and provide personalized learning advice and guidance.

Zou, S built an English teaching platform based on artificial intelligence technology and found that it significantly improved students’ English performance based on a follow-up survey [22]. Zhang, X et al. conceptualized an actual strategy for an intelligent classroom in English language and literature with artificial intelligence technology as the core logic, which has positive significance for the improvement of students’ overall English language proficiency [23]. Shin, M. H explored and dissected an AI technology-enabled English language teaching classroom model, and by analyzing student feedback, it was noted that this English language smart teaching classroom stimulated students’ self-efficacy [24]. Wu, S and Wang, F have demonstrated that the listening and speaking teaching mode of English majors with artificial intelligence technology as the underlying logic has promoted students’ self-management learning ability to a certain extent by conducting experimental English teaching classes [25]. Cuiye, Y et al. illustrated that the English teaching system based on AI technology creates a highly interactive and contextualized English teaching environment, which promotes teachers’ teaching and students’ English learning [26]. Tolstykh, O. M and Oshchepkova, T demonstrated AI technology-enabled English classroom teaching tools and launched an evaluation, which provides teachers with a lesson in the selection of intelligent tools for teaching English [27]. The research around AI technology-enabled English teaching mainly includes teaching models, teaching tool selection, and teaching effect evaluation, but the research is too much English teaching theory learning, ignoring the English application and practice.

In this paper, we first analyze the basic principles of using the Attention Knowledge Tracking model to predict students’ performance in exercises and to detect students’ knowledge mastery, and introduce two kinds of intelligent teaching-assistance systems at the topic level and at the skill level. An intelligent English teaching aid system based on knowledge tracking is proposed. The system model embeds a representation of English practice records and uses the sine-cosine function for position encoding. The gate mechanism is used to optimize the attention mechanism, and scaled dot product computation of attention is employed for multi-head attention operations. Based on the knowledge tracking framework, the intelligent English teaching and learning system is designed, which covers modules such as management, data management, classroom service, and examination service, and helps to better serve the integration of industry and education. The effect of the knowledge tracking model is tested through comparative experiments, and the application of the new teaching method in the senior English classroom is examined.

2

Fundamentals of Attention Knowledge Tracking and Intelligent Teaching and Learning System (ATKTS)

2.1

Attentional Knowledge Tracking Model

Attentional Knowledge Tracking (AKT) consists of four components: an encoder that collects questions, an encoder that collects knowledge, an attention-based knowledge decoder and a forward response prediction model. Based on the questions that students have answered, the question encoder for each question can generate context-based features. In the knowledge decoder, contextualized features are generated for the knowledge contained in each question based on the questions that have been answered by the students. The knowledge decoder utilizes the attention mechanism retrieval to obtain knowledge related to the past solved questions and the current question. The forward response prediction model utilizes the currently retrieved knowledge to predict the student’s performance on the exercises.

The input to the question collection encoder in the AKT model is the embedding vector $X = {x_{1}, x_{2}, \dots, x_{t}}$ $$X = \left\{ {{x_1},{x_2}, \ldots ,{x_t}} \right\}$$ and decomposes it into the query vector $Q = {q_{1}, q_{2}, \dots, q_{t}}$ $$Q = \left\{ {{q_1},{q_2}, \ldots ,{q_t}} \right\}$$ and the key vector $K = {k_{1}, k_{2}, \dots, k_{t}}$ $$K = \left\{ {{k_1},{k_2}, \ldots ,{k_t}} \right\}$$. After the monotonic attention mechanism the embedding vector X is transformed into $\bar{X} = {{\bar{x}}_{1}, {\bar{x}}_{2}, \dots, {\bar{x}}_{t}}$ $$\bar X = \left\{ {{{\bar x}_1},{{\bar x}_2}, \ldots ,{{\bar x}_t}} \right\}$$. $\bar{X}$ $$\bar X$$ is taken as one of the inputs to the knowledge collection decoder and decomposes it into the query vector $\bar{Q} = {{\bar{q}}_{1}, {\bar{q}}_{2}, \dots, {\bar{q}}_{t}}$ $$\bar Q = \left\{ {{{\bar q}_1},{{\bar q}_2}, \ldots ,{{\bar q}_t}} \right\}$$ and the key vector $\bar{K} = {{\bar{k}}_{1}, {\bar{k}}_{2}, \dots, {\bar{k}}_{t}}$ $$\bar K = \left\{ {{{\bar k}_1},{{\bar k}_2}, \ldots ,{{\bar k}_t}} \right\}$$. Similarly, the input to the knowledge collection encoder is the embedding vector $Y = {y_{1}, y_{2}, \dots, y_{t}}$ $$Y = \left\{ {{y_1},{y_2}, \ldots ,{y_t}} \right\}$$, which is transformed into the value vector $V = {v_{1}, v_{2}, \dots, v_{t}}$ $$V = \left\{ {{v_1},{v_2}, \ldots ,{v_t}} \right\}$$. After monotonic attention mechanism it is transformed into $\bar{Y} = {{\bar{y}}_{1}, {\bar{y}}_{2}, \dots, {\bar{y}}_{t}}$ $$\bar Y = \left\{ {{{\bar y}_1},{{\bar y}_2}, \ldots ,{{\bar y}_t}} \right\}$$. $\bar{Y}$ $$\bar Y$$ is used as one of the inputs to the Collected Knowledge Decoder and transformed into value vector $\bar{V} = {{\bar{v}}_{1}, {\bar{v}}_{2}, \dots, {\bar{v}}_{t}}$ $$\bar V = \left\{ {{{\bar v}_1},{{\bar v}_2}, \ldots ,{{\bar v}_t}} \right\}$$.

Each encoder and knowledge decoder has a query, key, and value embedding layer with dimensions D_q, D_k, and D_v in that order. In the encoder and knowledge decoder, AKT takes into account the effect of time on the decay of knowledge mastery, and introduces a time decay parameter to derive the S_t,τ. Next, the correlation between query vectors and key vectors is computed using the vector dot product, and then a Softmax activation function is used to compute the Attention Score α_t,τ: (1) $s_{t, τ} = \frac{\exp (- d (t, τ) \cdot q_{t}^{T} k_{τ})}{\sqrt{D_{k}}}$ $${s_{t,\tau }} = \frac{{\exp \left( { - d(t,\tau ) \cdot q_t^T{k_\tau }} \right)}}{{\sqrt {{D_k}} }}$$ (2) $α_{t, τ} = S o f t \max (s_{t, τ}) = \frac{\exp (s_{t, τ})}{\sum_{1 \leq i \leq t} \exp (s_{t, τ})}$ $${\alpha_{t,\tau }} = Soft\max \left( {{s_{t,\tau }}} \right) = \frac{{\exp \left( {{s_{t,\tau }}} \right)}}{{\sum\limits_{1 \leq i \leq t} {\exp } \left( {{s_{t,\tau }}} \right)}}$$

where attention score α_t,τ ∈ R^Dv×1, question vector q_t ∈ R^Dq×1, and knowledge vector k_T ∈ R^Dk×1. τ is the historical timestamp, and the metric value of the time difference in answering the exercises d(t, τ) = |t − τ|.

In the knowledge decoder, the AKT model makes interactions between the exercises completed at different times and the answer feedback value vector to predict the student’s performance on the pending exercises $\bar{r_{t}}$ $$\overline {{r_t}}$$, which is calculated as: (3) ${\bar{r}}_{t} = \sum_{t} α_{t, t} {\bar{v}}_{t}$ $${\bar r_t} = \sum\limits_t {{\alpha_{t,t}}} {\bar v_t}$$

The loss function of the AKT model is calculated as [28]: (4) $L = - \sum_{t} (r_{t} \log {\bar{r}}_{t} + (1 - r_{t}) \log (1 - {\bar{r}}_{t}))$ $$L = - \sum\limits_t {\left( {{r_t}\log {{\bar r}_t} + \left( {1 - {r_t}} \right)\log \left( {1 - {{\bar r}_t}} \right)} \right)}$$

where vector $r t \in {0, 1}$ $$rt \in \left\{ {0,1} \right\}$$ represents the true performance of the students’ answers to the exercises.

2.2

Intelligent Teaching Aid System

2.2.1

Topic-level intelligent tutoring system

Intelligent tutoring systems can be categorized into two types, which are topic-level and skill-level tutoring systems [29]. On the topic-level intelligent tutoring systems, students answer the questions directly. If the topic is answered correctly or incorrectly, all the skill points contained in the topic are also answered correctly or incorrectly. Thus, if a student answers skill point q_t correctly or incorrectly, he must answer the skill points in the set of associated skill points S_t correctly or incorrectly, since q_t and S_t are from the same topic. Suppose a student’s record of an answer in a topic-level intelligent tutoring system Topic q₁ consists of Skill s₁ and Skill s₃, and Skill s₁ and Skill s₃ are related skill points to each other. If a student answers Question q₁ incorrectly, then it also means that he has answered Skills s₁ and s₃, which make up Question q₁, incorrectly. Similarly, if a student answers question q₂ correctly, and question q₂ consists of skill s₅ and skill s₆, and skill s₅ and skill s₆ are related skill points, then it also means that he has answered skill s₅ and s₆ correctly, which make up question q₂. Therefore, the following relationship can be derived: (5) $a_{t} \equiv c_{t}$ $${a_t} \equiv {c_t}$$

That is, it means that in the topic-level intelligent tutoring system, at moment t, the right and wrong information of the student’s answering skill point q_t is consistent with the right and wrong information of the answering associated skill point set S_t.

2.2.2

Intelligent tutoring system at the skill level

The answer situation in a skill-level smart tutoring system is much more complex than in a topic-level smart tutoring system, assuming that there is a skill point KC₁, and its associated skill points are KC₂ and KC₃. A student can answer a skill point in a topic individually, and can answer that skill point one or more times. Therefore, if a student correctly answers skill point KC₁, it does not mean that the student can correctly answer KC₂ or KC₃.

On the surface, there is no apparent relationship of correctness of response between Skill Point q_t and the skill points in Linked Skill Point Set S_t. For example, a student’s answer to Question q₃₃ consists of Skill s₁₁ and Skill s₂₁, and Skill s₁₁ and Skill s₂₁ are related skill points. The student repeatedly answers question s₁₁ incorrectly, and even if he gets it right on the fourth time he answers question s₁₁, it shows that his mastery of skill point s₁₁ is poor, and similarly, his mastery of skill point s₂₁, which is an associated skill point of s₁₁, is poor, so the student, at time step 5, answers question s₂₁ incorrectly. The student’s answer data shows that if a student answers skill point q_t incorrectly multiple times, even if he finally answers skill point q_t correctly, it can indicate that he has a poor grasp of skill point q_t, and at the same time, his mastery of skill points in the associated skill point set S_t of q_t is not good, and it is very likely that he will answer the skill point in the associated skill point set S_t incorrectly. This can be expressed by a probability formula as: (6) $P (q_{t}) \approx P (S_{t})$ $$P\left( {{q_t}} \right) \approx P\left( {{S_t}} \right)$$

Where, $P (\cdot)$ $$P\left( \cdot \right)$$ indicates the student’s knowledge mastery of the skill point, and the larger the value indicates the more proficient knowledge mastery, $P (\cdot) \in (0, 1)$ $$P\left( \cdot \right) \in \left( {0,1} \right)$$.

According to the probability formula (6), it can be introduced: (7) $a_{t} \approx c_{t}$ $${a_t} \approx {c_t}$$

That is, it means that at the moment t, the student answers skill point q_t and the set of associated skill points S_t similarly, that is, if the student answers skill point q_t correctly, then he is also likely to answer the skill point in the set of associated skill points S_t correctly, and vice versa. According to Equation (7), the relationship between a_t and c_t can be introduced.

The students’ actual answers on the topic-level intelligent tutoring system and the skill-level intelligent tutoring system provide support for the knowledge tracking model of associated skills proposed in this paper.

3

Design of an Intelligent English Teaching and Learning System Based on Knowledge Point Tracking

3.1

Overall framework of the model

The model proposed in this paper is divided into four main parts: (1) An embedded representation module for the input of the exercise record. (2) Encoder based on self-attention mechanism. (3) Decoder based on self-attention mechanism. (4) Prediction result representation module. The focus of the model is on the encoder-decoder part, which are mainly implemented by the combination of multi-head attention computing module, feed-forward neural network, and layer normalization layer.

3.1.1

Embedded Representation of Exercise Records

The model proposed in this paper uses three embedding representation vectors. Let the currently processed student be i. For the k questions under the student ID, the three vector groups are vector group E_i converted from the sequence of question IDs, vector group S_i converted from the sequence of skill IDs, and vector group R_i converted from the sequence of answers, and there are k vectors in each group. The dimensions of each type of vectors need to be predefined, where the dimension of vector group R_i should be the sum of the dimensions of vectors E_i and S_i, taking into account the subsequent gating operations.

An embedded representation layer with randomly initialized parameters is used. In particular, for the vectors in the answer sequence vector set R_i, a normally distributed parameter initialization method with different expected values is set according to whether the answer is correct or not, i.e., μ > 0 when the answer is correct and μ < 0 when the answer is not correct. In addition, it is taken into account that the gradual adjustment of the embedded representation portion of the results in the backpropagation stage may result in a shift in the mean of the parameters as a whole, i.e., the mean of the results that are in fact answered correctly is gradually shifted to a negative value, and the mean of the embedded representation layer of the results that are answered wrongly is gradually shifted to a positive value. The Leaky-Relu was used as the activation function for the error conduction calculation in the backpropagation stage, as the mean values of the embedded representation layer of the results of the correct answers gradually shifted to positive values. This ultimately serves to avoid or delay such anomalous shifts.

As inputs to the subsequent attention module, E_i and S_i perform the vector splicing operation according to the correspondence between the exercise ID and the skill ID. The length of the final expanded vector is equal to the length of the vector in vector group R_i.

3.1.2

Location coding

In this paper, absolute position coding, i.e., sine-cosine function position coding, is used for the position coding of the proposed model [30]. The sine-cosine function position encoding is not a single value, but a d-dimensional vector containing information about a specific position in the sequence. This encoding is not included in the training parameters of the model, but is used to give each element information about its position in the sequence. That is, the model input is enhanced by injecting information about the order of the elements. For length d and position pos in the sequence, the value of dimension i in the position encoding vector is: (8) $\begin{matrix} P E_{p o s, 2 i} = \sin (p o s / 10000^{2 i / d}) \\ P E_{p o s, 2 i + 1} = \cos (p o s / 10000^{2 i / d}) \end{matrix}$ $$\begin{array}{rcl} P{E_{pos,2i}} = \sin \left( {pos/{{10000}^{2i/d}}} \right) \\ P{E_{pos,2i + 1}} = \cos \left( {pos/{{10000}^{2i/d}}} \right) \\ \end{array}$$

A series of values form the position encoding vector, where $i \in {0, 1, 2 \dots d_{\mod e l} / 2}$ $$i \in \left\{ {0,1,2 \ldots {d_{\bmod el}}/2} \right\}$$. Generates the position encoding and then combines it with the embedding representation vector in an additive manner.

3.1.3

Resulting control gates

In order to optimize the computational effect under the embedded representation as well as the dot product attention mechanism, this model uses a gate mechanism instead of a simple extension, i.e., for the inputs of the value $(V)$ $$\left( V \right)$$ part of the multi-head attention module, there is: (9) $E_{i}^{v} = (E_{e_{i}, s_{i}} + P_{i}) ⊙ \tanh (E_{r_{i}})$ $$E_i^v = \left( {{E_{{e_i},{s_i}}} + {P_i}} \right) \odot \tanh \left( {{E_{{r_i}}}} \right)$$

where $E_{e_{i}, s_{i}}$ $${E_{{e_i},{s_i}}}$$ is the embedding representation vector of the spliced topic and skill information, P_i is the location encoding, $E_{r_{i}}$ $${E_{{r_i}}}$$ is the embedding representation of the result, and ⊙ is the Hadamard product operation. Unlike the embedded representation of topic ID and skill ID, $E_{r_{i}}$ $${E_{{r_i}}}$$ does not need to be combined with positional encoding.

3.1.4

Multi-attention operations

The multi-head attention module is the core part of the encoder and decoder. This model uses a scaled dot product to compute the attention, i.e: (10) $A t t e n t i o n (Q, K, V) = s o f t \max (\frac{Q K^{T}}{\sqrt{d_{k}}}) V$ $$Attention\left( {Q,K,V} \right) = soft\max \left( {\frac{{Q{K^T}}}{{\sqrt {{d_k}} }}} \right)V$$

Usually for queries in self-attention operations, the keys and values are denoted by Q, K and V, which come from the embedded representation vectors of topic information after performing linear operations [31]. In the model proposed in this paper both Q and K come from the vectors obtained by further processing of the topic ID and skill ID, while V comes from the result obtained after the former is gated with the result vector. Namely: (11) $\begin{array}{rcl} Q_{i} & = & W_{q} (E_{e_{i}, s_{i}} + P_{i}) \\ K_{i} & = & W_{k} (E_{e_{i}, s_{i}} + P_{i}) \\ V_{i} & = & W_{v} E_{i}^{v} \end{array}$ $$\begin{array}{rcl} {Q_i} &=& {W_q}\left( {{E_{{e_i},{s_i}}} + {P_i}} \right) \\ {K_i} &=& {W_k}\left( {{E_{{e_i},{s_i}}} + {P_i}} \right) \\ {V_i} &=& {W_v}E_i^v \\ \end{array}$$

Where W is the linear transformation matrix, this operation is the operation that divides the heads in the multi-head attention mechanism. After this operation, different representations can be obtained from the original query, keys and values, and finally the results of the different attention operations for each head are then spliced together before inputting to the feed-forward neural network. This model uses 8 different sets of W to divide the input section into 8 heads to capture richer features mapped to multiple linear spaces. The self-attention mechanism is used to compute the attention score, which is multiplied in the form of coefficients with the vectors that have been subjected to gate transformation and linear operations, and the heads are spliced and merged, ultimately converting the input vectors of individual elements into a representation in the form of groups of vectors for the elements in the sequence.

For decoders, there is a difference from encoders. The second stage of the decoder’s multi-head attention receives not only the input from the previous stage, but also the input from the encoder. The inputs from the encoder are set to ends K and V of the self-attention mechanism of the stage, while the output of the first stage of the decoder is set to end Q of the inputs of the second stage. This setup allows each position of the decoder to perform an attention operation on each element of the input sequence.

3.1.5

Model outputs and other structures

A two-layer feed-forward neural network is used in the model with a linear revision unit (ReLU) as the activation function: (12) $F F N (x) = \max (0, x W_{1} + b_{1}) W_{2} + b_{2}$ $$FFN(x) = \max \left( {0,x{W_1} + {b_1}} \right){W_2} + {b_2}$$

Considering that the model stacks two layers of the multi-head self-attention module, and that in networks with deeper layers, residual connections optimize gradient conduction during back propagation and enhance the representation of the deep network, the model uses residual connections in the value processing part of the multi-head self-attention module as well as in the feedforward neural network part. Layer normalization is also used for the output of the attention module as well as the feedforward neural network. Normalization between layers stabilizes the gradient of the neural network, eliminating the effects of vanishing or exploding gradients, and also speeds up the convergence of the training.

Eventually, the output of the decoder goes through the fully connected layers and undergoes a Sigmoid compression function to finally output the correctness of the prediction of the decoder’s input topic: (13) $p_{i} = S i g m o i d (x_{i} W + b)$ $${p_i} = Sigmoid\left( {{x_i}W + b} \right)$$

In the training phase of the model, the optimization objective is to minimize the binary cross entropy loss. The prediction values are taken in the range of $[0, 1]$ $$\left[ {0,1} \right]$$ and the dataset labels are taken as ${0, 1}$ $$\left\{ {0,1} \right\}$$.

3.2

Intelligent English Teaching and Learning System Design

Considering the good performance of intelligent assistants in online learning sites, it makes sense to integrate such intelligent assistants into offline teaching. By implementing a system with intelligent teaching assistant functions, teacher’s teaching and student’s learning can be observed and recorded at a fine-grained level, personalized teaching is supported by accurate student models, students can easily understand their own learning status, and teachers can get teaching planning after multi-objective and multi-factor teaching optimization, and so on.

Based on the above objectives, the functional requirements of the system include system management function module, course data management function module, classroom service function module, examination service function module, student model function module and recommendation function module, in order to realize the management of the relevant data involved in accessing the users and the courses, the collection and provision of classroom interaction information and examination information, the construction and updating of the student model, as well as the recommendation of the students’ learning and teachers’ teaching. Recommendations for students’ learning and teachers’ teaching. 1)

The system management functions include user management, permission management and log management: user management means the registration, login and account management of teachers and students. Privilege management is responsible for managing the privileges of each account, e.g. teachers can only manage the courses they teach. Log management, on the other hand, is responsible for recording key operations performed by all users, such as logging in, changing passwords and initiating polls.

2)

Considering that the target users of the system are teachers and students learning in the classroom, the registration function of general users is restricted. The system administrator account has the highest authority and cannot be created or deleted. Teacher accounts are created and assigned by the system administrator. The administrator can assign or disable the privileges of a teacher account to manage the courses and classes he/she teaches, as needed. Teacher accounts can also be reset through the administrator’s account if their password is lost. Student accounts require the account ID to be bound to the student number in the list of students in the course entered by the teacher beforehand during registration. After completing the basic information registration, student accounts are required to complete a personality test in the form of a questionnaire, which serves as the basis for the student model. Student accounts can change their own passwords, and if they lose their passwords, they can be reset by the course instructor or system administrator.

3)

The course data management function includes course management and knowledge point management: course management is responsible for recording the basic information of the course, such as class hours, instructors, number of students and syllabus. Knowledge point management is responsible for the content management of the course specific to the granularity of each knowledge point, including knowledge point ID, name, difficulty, expected teaching time, supporting teaching materials and videos.

Among them, the basic information of the course is entered by the system administrator, who can only arrange teachers and assign permissions for the courses that have been entered. The knowledge point information in the knowledge point management is only entered and modified by teachers with knowledge point management privileges, while the supporting materials and videos can be accessed, added or deleted by all teachers teaching the subject.

4)

Classroom services include classroom planning, classroom interaction and after-school homework: classroom planning refers to a variety of recommended plans prepared by the system for teachers or students, according to which users can teach, study or modify their plans. Classroom interaction includes the teacher initiating classroom polls or classroom quizzes, and students answering and viewing the results. After-class assignments are responsible for the teacher’s posting of assignments and the students’ viewing of assignments.

Among the sub-functions in the classroom service module, the classroom planning function can be divided into individual learning recommendation, group teaching recommendation and group recommendation. Student users can view the personal learning knowledge point recommendations that the system pushes for them and can make adjustments based on them. In addition, students can also view the teaching arrangements issued by teachers. Teacher users can view the student classroom grouping plan and group teaching plan pushed by the system, and can be modified on this basis.

The details of the functions are as follows: the system reminds the teacher when the teaching time is approaching according to the pre-defined teaching plan. Teachers can play the materials or videos pre-stored in the system. After the teaching or video playback is completed, the teacher initiates a quiz or classroom poll, students participate in the quiz or poll, and the system saves the students’ answers or choices and updates the student model.

5)

The student modeling module includes student information management, student modeling and knowledge tracking: student information management is responsible for managing students’ basic information such as name, student number and class. Student modeling is responsible for building a student model based on students’ learning interaction data and basic information, including students’ mastery of each knowledge point. Knowledge tracking is responsible for updating and tracking the student model at any time during the learning process.

6)

Examination service module includes examination content management and examination result management: examination content management is responsible for the information management of the content and score of the examination. Examination results management is responsible for the entry and viewing of students’ examination results. After each score entry, the system will update the student model.

7)

Recommendation function module is divided into student learning recommendation, teacher teaching plan recommendation and group recommendation: student learning recommendation for the system according to the student’s learning situation for each student to recommend the most suitable learning knowledge path. Teacher teaching program recommendation is a pointer to the best teaching arrangements for the entire class, including the expected teaching of knowledge points in each lesson and the recommended teaching time for each knowledge point. Grouping recommendation means that the system automatically sends the best grouping plan, i.e. the list of members of each group, to the teacher when working or teaching in groups. All three recommendation functions will automatically invoke their respective recommendation algorithms to update the plan and push it to the teacher or students after each learning activity or student model update.

The grouping recommendation combines students’ personality traits and learning situations to maximize the cooperation and complementarity of students in the group, which is implemented by genetic algorithms. Student learning recommendation and teacher teaching plan recommendation are based on reinforcement learning algorithms, which train their respective recommendation models by inputting students’ learning trajectories as well as related information to give the most suitable learning or teaching methods, in order to ensure the effectiveness of the recommended content and efficiently promote students’ learning.

4

Experimental design

4.1

Training and Testing of Knowledge Tracking Models

In order to validate the effectiveness of the method proposed in this paper, this section chooses to train and test the knowledge tracking model that incorporates learning features and topic difficulty on the ASSISTments2009 dataset. In this paper, the learner data in the dataset is randomly divided, with 70% of the data constituting the training set and 30% of the data constituting the testing set. The evaluation metrics of the algorithm in this paper are AUC and Precision, and corresponding comparison experiments are designed to test the effectiveness of the model.

4.1.1

Comparative experiments

1)

Model selection and parameter setting

The knowledge tracking model in this paper is compared with the existing mainstream knowledge tracking model to verify the effectiveness of the proposed method. The comparison models are DKT, DKT-F, DKVMN, and DKVMN-LA. The hidden layer nodes of DKT were set to 200, the batch_size was set to 64, the learning rate learning_rate was 0.001, and the optimizer in training used Adam. The DKT-F, DKVMN, and DKVMN-LA models all use Adam as the optimizer with a learning rate of 0.001. The size of the hidden layer of the GRU neural network for the model in this paper is 200, the number of training rounds epoch is 100, batch_size is 64, and the learning rate is set to 0.001.

2)

Analysis of results (1)

Importance of learner features

In this section, LightGBM algorithm is used to realize the prediction of learner performance by analyzing multiple features in the dataset, and obtain the importance of each feature in the dataset to the learner’s answering situation, which provides the basis for subsequent feature selection. Experimentally selected features with higher importance are encoded with other features to compose the input of the model in this paper.

The importance ranking of learning features obtained after four-fold cross-validation of LightGBM algorithm’s prediction of learning performance is shown in Figure 1. As can be seen from the figure, the last three features have the lowest importance ranking, so the remaining features with high importance are utilized as the learner feature set in the subsequent deep knowledge tracking model training.

(2)

Analysis of the results of comparative experiments

The comparison results of the performance of this paper’s model with other knowledge tracking models on the ASSISTments2009 dataset are shown in Table 1. As obtained from the table, compared with the four models DKT, DKT-F, DKVMN, and DKVMN-LA, the overall performance of the model proposed in this paper outperforms the other in-depth knowledge tracking models, with an AUC value of 0.966, which is 14.18%, 9.40%, and 10.40% better than that of DKT, DKT-F, DKVMN, and DKVMN-LA, respectively, 5.23%, and its Precision value is also the highest among the models, which indicates that the method proposed in this section of the paper to add learner features to the model is correct and effective, and has substantially improved the comprehensive performance performance of the knowledge tracking model algorithms, as well as the prediction accuracy of the learners’ knowledge acquisition and future performance.

Table 1.

The model performance comparison

Model	AUC	Precision
DKT	0.846	0.811
DKT-F	0.883	0.837
DKVMN	0.875	0.831
DKVMN-LA	0.918	0.873
Ours	0.966	0.918

4.1.2

Visualization of learners’ knowledge states

This subsection randomly selects learners’ partial answer data in the ASSISTments2009 dataset, which contains 5 knowledge points and 16 time-steps of question interaction records, and uses this paper’s model to predict them and get the visualization results of learners’ knowledge level change with time-steps, as shown in Figure 2.

In Fig. 2, the horizontal axis represents the learner’s question-answering sequence, and the learner’s question interactions at each time step are represented using a binary group for the number of the question, corresponding to the answering situation, where 0 represents an incorrect answer and 1 represents a correct answer. The vertical axis on the left side of the graph is the name of the specific knowledge point corresponding to the topic. The change in the color of the grid in the graph represents the change in the degree of knowledge mastery, the closer the color is to the light color, the higher the corresponding degree of knowledge mastery of the learner, and conversely, the closer the color is to the dark color, the lower the corresponding degree of knowledge mastery of the learner.

First of all, in terms of the answers to question 35, the learners answered question 35 four times in total, the first and the third time they answered incorrectly, and the corresponding grid color is close to dark, indicating that the learners’ mastery of the knowledge of question 35 was low at that time step. The second and fourth times were answered correctly, and the color of the grid became lighter, indicating that the learners’ mastery of the knowledge points of Question 35 increased at that time step. Second, as a whole, the color of the grid tends to become lighter, indicating that the learners’ mastery of the knowledge points contained in the questions have increased after answering the questions several times.

It can be seen that the model in this paper can obtain the changes in the learners’ knowledge mastery according to the learners’ answering records, realize the dynamic tracking of the learners’ knowledge level, help the online platform to complete the accurate recommendation of the learners, help the learners to check the omissions and make up for the deficiencies, get rid of the problematic tactics, and further improve the personalized service.

4.2

The Application of Modeling in Higher Vocational English Classroom Teaching

The knowledge points of a school’s senior English course are sorted out and the knowledge units contained in the course are extracted. Each knowledge unit is taught as an independent teaching task and is scheduled to be completed in one teaching period, and homework or quizzes are assigned after the teaching of each knowledge unit is completed.

Two hundred students were selected from the teaching process in recent years and their completion of the assignments or quizzes for each unit, as well as other units of knowledge covered by the assignments or quizzes, were counted. When counting students’ mastery of what they have learned, they are considered to have learned the relevant knowledge if their scores on the assignments or quizzes are “good” or “good” or above, otherwise they are considered not to have learned the relevant knowledge. Based on the intrinsic connection between the knowledge points to establish the knowledge structure, and then use the parameter learning algorithm to get the dependence probability table between the knowledge points, take the probability term whose value is true in the probability table of each parent node as the strength of the association between this node and its child nodes, and then combine with the experience to get the knowledge structure map of the course with certain corrections.

Taking the knowledge of senior English as the basic knowledge of the course, the degree of a student’s mastery of the basic knowledge of the English course = this student’s grade in the English course/100. Figures 3 and 4 compare the pass rates of each knowledge point of the original teaching process and the first pass rates of each knowledge point of the knowledge-tracking-based teaching process, and the pass rates of each knowledge point of the original teaching process and the final pass rates of each knowledge point of the knowledge-tracking-based teaching process, respectively.

The results showed that the first-time pass rate of the teaching method incorporating knowledge tracking was not much different from that of the original teaching method, and the average pass rates of the two were 79.3% and 78.4%, respectively. And after using the knowledge-tracking teaching method for long-term learning, the English course pass rate of senior students changed and the average pass rate reached 89.4%, which was about 10.1% and 11% higher than the first-time pass rate and the original teaching method, respectively.

5

Conclusion

Under the guidance of the concept of industry-teaching integration, this paper designs a knowledge point tracking framework using the attention knowledge tracking model, and based on it, constructs an intelligent teaching assistance system for higher vocational English classrooms. On the ASSISTments2009 dataset, the AUC of this paper’s model = 0.966, which is 14.18%, 9.40%, 10.40% and 5.23% higher than that of the DKT, DKT-F, DKVMN, and DKVMN-LA models, respectively. The accuracy reaches 0.918, which is also better than the above comparison models, verifying the accuracy of the knowledge tracking of the method in this paper. After the introduction of knowledge tracking teaching in a school’s senior English course, its course pass rate increased from 79.3% to to 89.4%, proving that the method can improve the teaching effect of senior English teaching.

Language:: English

Publication timeframe:: 1 times per year
Journal Subjects:: Life Sciences, Life Sciences, other, Mathematics, Applied Mathematics, General Mathematics, Physics, Physics, other

Journal RSS Feed

Optimization Design of Artificial Intelligence Aids in Higher Vocational English Classroom Teaching under the Concept of Industry-Teaching Integration

Tingting He

Jiguo Yao

Yonghui Lai

Yuhan Zou

Published Online: Sep 29, 2025

Received: Dec 20, 2024

Accepted: Apr 16, 2025

DOI: https://doi.org/10.2478/amns-2025-1102

KeywordsKnowledge point tracking, Artificial intelligence, Intelligent teaching aid system, Senior vocational English classroom teaching, Industry-teaching integration

© 2025 Tingting He et al., published by Sciendo

This work is licensed under the Creative Commons Attribution 4.0 International License.

Keywords
Knowledge point tracking, Artificial intelligence, Intelligent teaching aid system, Senior vocational English classroom teaching, Industry-teaching integration