Optimisation of Teaching Methods and Practical Exploration of Teachers’ Teaching Methods in Vocational Education Based on AI Assistance

Bidirectional long and short-term memory neural network (BILSTM), as an extension of the RNN model [19], is suitable for sequence modelling of textual information. This model is based on the introduction of bi-directional thinking to the Long Short-Term Memory Neural Network (LSTM), which encodes both front-to-back as well as back-to-front information, making full use of contextual information and is more accurate than the unidirectional LSTM.

BILSTM is composed of a Forward LSTM and a Backward LSTM together, and both the Forward and Backward LSTMs are connected to the Output layer, i.e., the final output of the Output layer is the result of splicing the implicit layer vectors of the LSTMs of the two directions, and the rest of the transfer process and the working principle are the same as that of the LSTM model.

The forgetting gate, input gate, and output gate are the key points of the long and short-term memory neural network model. LSTM controls the cell state by processing the information flowing into the model through these three gating units so that the information can pass through selectively: the forgetting gate is used to decide which part of the information is to be discarded. The input gate determines which information to leave in the cell state, and the output gate determines which information to output.

Oblivion gate is the first step of LSTM model, by reading h^(t−1) and sample input x^(t) and superimposing the two, using sigmoid activation function to get f^(t) and f^(t) ∈ [0,1], f^(t) is zero means all the previous layer cellstate will be rounded off, f^(t) is 1 means all left behind, this process is shown in equation (1): 1 f(t)=σ(Wth(t−1)+Utx(t)+bt)

The input gate is mainly divided into two parts, the first part firstly h^|t−1| and x^(t) are passed through the sigmoid activation function to obtain i^(t), and i^(t) ∈ [0.1], and then the tanh function is used to obtain a^(t), the computation process is shown as follows: 2 i(t)=σ(Wth(t−1)+Utx(t)+bt) 3 a(t)=tanh(Wah(t−1)+Uax(t)+ba)

The second part of the input gate makes Hadamard product for i^(t) and a^(t) based on f^(t),i^(t) and a^(t) calculated above and updates the cellstate c^(t−1) to c^(t). The update process is shown in equation (4): 4 c(t)=c(t−1)□f(t)+i(t)□a(t)

The final output value of the output gate is based on the cellstate, specifically, the output gate first superimposes h^(t−1) and x^(t) through the sigmoid activation function, and then multiplies it with the result of c^(t) that has gone through the tanh function to finally determine the output value of h^(t). The computation process is shown below: 5 o(t)=σ(Woh(t−1)+Uox(t)+bo) 6 h(t)=o(t)□tanh(c(t))

Transfer learning refers to learning in the target domain (representing the new domain to be learned) by applying the knowledge of the source domain (representing the knowledge that has already been learned in a certain domain using a method or model) through the use of the idea of transfer, the use of this method has a requirement on the target domain: it must be similar to the source domain, and then through the rational use of this similarity to share the information of the knowledge structure, the parameters in the model, etc., so as to get better results in the target domain tasks to get better results.

Among them, the model-based migration method is widely used, which utilizes shared parameters between the source and target domains to achieve the target task.

BERT is a multilayer Transformer bidirectional encoder released by Google [20], which belongs to a type of word vector model. Deep bidirectional representations are pre-trained by this model by jointly tuning the left and right contexts in all layers, and it has excellent performance on multiple NLP domain tasks.

The model is divided into two main parts: pre-training and fine-tuning. The first part of it is trained on unlabelled data for different tasks; the second part first initialises the model with the parameters trained in the first part and then utilises the labelled data from subsequent NLP tasks (e.g., question-answer, inference, named entity recognition, etc.) during the fine-tuning of the model. Meanwhile, the model can have three types of inputs, namely text vectors, word vectors and memory location vectors, and the activation function uses gel (Gaussian Error Linear Unit), which is an improvement of relu by adding statistical features to it, and the mathematical expression is shown in Eq. (7): 7 GELU(x)=xP(X<=x)=xΦ(x) Where Φ(x) refers to the probability function of the normal distribution.

Incorporating a knowledge graph in the recommendation method of vocational education learning resources can not only further improve the recommendation performance of vocational education learning resources but also enhance the interpretability of the method. The current techniques for combining knowledge graphs with recommendation domains can be mainly classified into two categories, and the classification results are shown in Figure 1.8 (k,T2,T2,T2)=−12

Figure 1.

A recommendation based on the knowledge map

Embedding-based recommendation methods adopt the main idea of using knowledge graphs to get embedded representations of nodes and edges in low-dimensional space vectors to enrich the representations of students and learning resources. However, this kind of method often does not make use of the correlation information between students and learning resources and lacks reasoning ability and weak interpretability.

Path-based recommendation methods adopt the main idea of treating the knowledge graph as a heterogeneous mesh graph and, at the same time, consider the meta-path structure between students and learning resources so as to obtain the hidden relationship information in the graph and improve the performance of the related learning resources recommendation, with good interpretability. However, this type of approach is highly dependent on the meta-path and requires the participation of experts in the planning process of the graph.

There are many ways to compute the similarity between user interests and resources, among which content-based recommendation uses the cosine similarity of text as the similarity between user interests and resources. There are different ways of computing the similarity between user interests and learning resources in different recommendation methods.

After completing the word vector representation of the resource, it is used to calculate the similarity of the text as well as the similarity between the user’s interest and the resource. The similarity is defined as follows: 9 Interest(u,i)=1|N(u)|*∑jeN(u)w→i·w→j| w→i |·| w→j | U represents the set of users, u represents the target user, N(u) represents the set of historical resources of this user, |N(u)| represents the number of the set of historical resources of the user, and w→i , w→j . represents the word vector representation of resource i and resource j respectively. In the similarity calculation, the higher the textual similarity between the resource and the user’s historical resources, indicating that the more similar the resource is to the user’s preference, the more recommendable the learning resource is.

In this paper, due to the limitation of the dataset, we use textual similarity to represent the similarity between user interests and resources. With more data, this section can also use more features to model user interests and use other methods, such as LFM, LDA topic model etc., to compute user preferences.

Knowledge connectivity is used to measure the close relationship between knowledge points. In the recommendation problem of teaching resources, the closer the subject knowledge points contained in a teaching resource are connected to the student’s cognitive knowledge points, the more the resource is considered to meet the student’s learning needs, and the more it is worth recommending.

For any two knowledge points k1 k2, their shortest path distance is the shortest path distance that can be retrieved from the knowledge graph. The shortest path distance between knowledge points ShortestPath(k2, k1) formula is defined as follows: 10 ShortestPath(k2,k1)={ ShortestPath (k2,k1)1 ( if k2∈KSi and k1∈KSi)0 ( if k2=k1 or k2 There is no path between

Based on the definition of the shortest path distance between knowledge points, the formula for the degree of knowledge connectivity between Knowledge Point k2 and Knowledge Point k1 is defined as follows: 11 KCD(k2,k1)={ in |k2|ShortestPath(k2,k1)0 (if ShortestPath(k2,k1)=0)

The above formula indicates that when the shortest path distance from knowledge point k2 to k1 is not 0, the knowledge connectivity from knowledge point k2 to k1 is the entry degree of knowledge point k2 divided by the shortest path distance from k2 to k1. ShortestPath(k2, k1). When the shortest path distance from knowledge point k2 to k1 is 0, it means that there is no path between knowledge points k1 and k2 or they are the same knowledge point, and it is considered that there is no relationship between the two knowledge points when there is no path between them. When there is no path between the knowledge points, it is considered that there is no relationship between the two knowledge points, and when it is the same knowledge point, it has no value for the recommendation of the learner, so its knowledge connectivity is set to 0.

In the knowledge graph G(V,E), V represents the collection of knowledge and E represents the collection of relationships between knowledge points. For the target user u who has had historical behaviour, its user knowledge base K(u) is the set of all knowledge points contained in the user’s historical resources, and the resource knowledge base K(S) contains the set of all knowledge points of the resource S. The knowledge connectivity from resource S to user knowledge base K(U) also represents the knowledge connectivity between resource S and target user u. The formula is expressed as follows: 12 KCD(S,u)=KCD(S,K(U))=∑k,eK(S)in| ki |ShortestPath(ki,K(U))

In the above formula, k_i denotes any knowledge point in the resource knowledge base K(S), and the in-degree of a knowledge point in |k_i| represents the importance of the knowledge point. We believe that the larger the in-degree of a knowledge point is, the more important it is, and the more it is worth learning and recommending. ShortestPath(k_i,K(U)) represents the shortest path between knowledge point k_i and all the knowledge points in user knowledge base K(U), the shorter the path, the tighter the knowledge connection.

The formula describes the meaning of: the shorter the path between knowledge points, the tighter the knowledge connection, the greater the degree of knowledge into the knowledge point, the more important the knowledge point. Combining the two, the greater the knowledge connectivity, the closer the connection between the knowledge sample of the resource and the user’s knowledge base, the more suitable the resource is for recommending to the user to learn.

When recommending learning resources, not only the learner’s interest is taken into account, but also the connectivity between the resource’s knowledge and the user’s knowledge level. According to the above definition of interest similarity and knowledge connectivity, the two are fused in a linearly weighted manner as the basis for resource recommendation. Parameter α is used to balance the weighting relationship between learner interest and cognitive level, so that the recommendation result does not lose the learner’s interest, but also meets the learner’s knowledge and cognitive level for the next step of learning. The degree of connectivity between learners and resources is defined as follows: 13 Sim(u,s)=(1−α)·Interest(u,s)+α·KCD(s,u)

The previous section gave a theoretical introduction to the idea of recommendation based on user interest and resource similarity and knowledge connectivity, and this section gives a detailed recommendation process on the basis of the above theory, and the recommendation process is shown in Figure 2.

Figure 2.

Recommendation system based on the knowledge map

The specific implementation steps of this recommendation method are as follows: 1)

The resource text data is subjected to lexical and de-duplication operations, and the keywords are extracted as the semantic features of the resources, and the keywords are represented as 50-dimensional word vectors, and the resource S_i is denoted as S_i = {w_i1, w_i2,…, w_in}.

The online teaching platform of a vocational school has been in operation for three years, providing sufficient data to support this personalized recommendation test. There are more than 100,000 resource data on the platform, covering multiple disciplines and resource types, as shown in Table 1. The network teaching cloud platform has 3500 student user records, and the data are divided into the training set, testing set and experimental set according to the ratio of 2:1:1, which are used for model training, method feasibility and validity testing, and Table 1 shows the resource profile of the network teaching cloud platform.

The user data of the online teaching cloud platform in the test set is randomly selected for analysis, and the method of this paper is applied to provide personalized resource recommendations. The behavioral records of the 10 selected cloud platform users are shown in Table 2.

According to the above data and user browsing content, the method of this paper is used to analyse the behavioural characteristics and recommend personalized resources, and the results of personalized recommendations are shown in Table 3. The method of this paper recommends different teaching resources for users, the degree of personalisation is good, and the matching degree between recommended content and students’ browsing content reaches more than 90%.

User surveys are mostly implemented by users completing questionnaires or recommending specific tasks on the system. In this study, we chose questionnaires that were completed by users. Their feelings can reflect the subjective evaluation of the system by users after using it, and the questionnaire designed for this study was also designed around this theme.

The questionnaire is designed to obtain learners’ subjective feelings about this recommender system in order to evaluate it. Its content is divided into three dimensions: interface ease of use, satisfaction with resource recommendations and overall experience of use, of which interface ease of use contains two questions with serial numbers 1 and 2, satisfaction with resource recommendations contains three questions with serial numbers 3, 4 and 5, and overall experience of use contains three questions with serial numbers 6, 7 and 8. 5 answers are used for the eight questions in this questionnaire. The level of agreement among learners with the questions was determined by calculating the proportion of people who agreed with the questions.

The target population of this questionnaire remained 100 learners in the school. The number of questionnaires distributed was 100, and the number of returned questionnaires was 100, with no invalid questionnaires. The statistical results of the questionnaire are displayed in Table 4. More than 90% of people chose to strongly agree on the eight questions designed, particularly when it comes to the overall feeling of use. The students chose to be very satisfied with the average number of people reached 95.33%, which illustrates that the intelligent recommendation system of teaching cloud resources based on the knowledge graph proposed in this paper has been affirmed by the majority of the people and that it has an effect of improving the professional performance of students in vocational schools.

In this study, two computer science major classes at a vocational school in the city of A, a province, with 50 students in each class, were selected as experimental subjects. In this experiment, the independent variable is the teaching methods used in the two classes, the dependent variable is the student’s academic performance, and the factors indispensable to teaching activities, such as the teaching environment and teaching facilities, belong to the irrelevant variables. When conducting teaching practice, attention should be paid to controlling the influence of irrelevant variables on the experiment. Therefore, except for the teaching method, the same teaching conditions should be used in both classes of the control group and the experimental group.

After the teaching practice was carried out in two classes of computer majors in this vocational school in city A, a stage test was organised for the students in order to assess the learning situation at this stage. The exam was closely focused on the learning content of the chapter. The exam was in the form of an offline paper version of the question paper, divided into subjective and objective questions, to comprehensively examine the students’ mastery of the knowledge they have learned and their ability to apply it. At the end of the quiz, the students’ answer sheets were reviewed, and the students’ scores were recorded in an Excel table for statistical analysis. The distribution of the scores of this quiz for the two classes is shown in Figure 3, with the top corner ordinal numbers 1-5 representing the five intervals of the students’ scores of the range of 91-100, 81-90, 71-80, 61-70, and less than 60, respectively.

Figure 3.

Student achievement

Comparing the results of the experimental class and the control class, the number of students in the experimental class in the high scoring range and the number of passing grades have increased to different degrees compared to the control class. In the experimental class, there were seven students in the 91-100 range, five more than in the control class. The number of students with scores between 81-90 is 20, which is the largest percentage of the class size at 40 percent. Fifteen students, or 30 percent, scored between 71 and 80. Another six students scored between 61-70, and only two students scored below 60. Overall, most of the students in the experimental class achieved good results after adding the optimal teaching method to the teaching process, and they were able to construct, internalize, and absorb knowledge actively.

In order to test whether the teaching method plays a positive role in student’s academic performance and to understand students’ learning feelings during the teaching process, a questionnaire was distributed to the students in the experimental class after the experiment, and their attitudes and views on this teaching mode were understood by analysing the students’ feedback on the questionnaire. The questionnaire was designed to cover four aspects: students’ mastery of knowledge, participation in the classroom, interest in the course, and satisfaction with the course. Two questions were designed for each aspect, using a five-point scale, and in terms of scoring rules, this questionnaire was set up with five options of Strongly Agree, Agree, Generally Agree, Disagree, and Strongly Disagree, which were assigned the values of 5, 4, 3, 2, and 1 respectively, with the larger values representing, the more the respondents agree with the statement. Table 5 displays the particulars of the questionnaire.

After the questionnaire data were sorted out, the frequency and proportion of each option in each question were counted, and the results were organized and visualized to analyze the students’ attitudes and opinions in the four aspects of “mastery of knowledge”, “participation in the classroom”, “interest in the course” and “satisfaction with the course”.

Figure 4 shows the statistics of the students on the question of “the classroom of optimized teaching method is beneficial to my mastery of course knowledge.” the serial numbers of vertical coordinates 1-5 indicate the five indicators of disagreeing, disagreeing, generally, agreeing, and agreeing very much, and 90% of the students chose “agreeing very much” or “agreeing”. 90% of the students chose “strongly agree” or “agree”, accounting for more than four-fifths of the students in this class, and most of the students feel that this teaching method is more conducive to their mastery of the knowledge of the course.

Figure 4.

Question 1 statistics

Figure 5 shows the statistics of students’ responses to the question “Optimized teaching methods in the classroom can drive me to learn the knowledge I need to use to complete the task”, and 88% of the students chose “Strongly Agree” or “Agree”. The number of students who chose “strongly agree” or “agree” was 88%, which means that this teaching model can promote students to actively learn the knowledge related to the task for the purpose of completing the task.

Figure 5.

Question 2 statistics

Figure 6 shows the results of the survey on “The use of the optimized teaching method has improved my participation in class”, more than half of the students chose “Strongly Agree”, and nearly 30% chose “Agree”. The majority of the students were able to actively participate in the classroom activities under this teaching model.

Figure 6.

Question 3 statistics

Figure 7 shows the results of the survey on “I can actively participate in the group work in class”, with 20 students selecting “strongly agree” and 20 students choosing “agree”. In addition, there was one student who chose “disagree,” and one student who “disagreed”, and most of the students actively participated in group work to explore ways to solve problems and complete tasks through group cooperation.

Figure 7.

Question 4 statistics

Figure 8 shows the results of the survey on “I am very interested in the content and teaching method of the course after using the optimized teaching method”, and the students with a positive attitude accounted for 96% of the total number of students in the experimental class, indicating that the classroom teaching model based on the optimized teaching method is attractive to most students and can increase students’ interest in the classroom to a certain extent.

Figure 8.

Question 5 statistics

Figure 9 shows the statistical results of “the use of optimized teaching method has stimulated my interest and enthusiasm in the C programming course” a total of 46 students agreed, accounting for 92% of the class, indicating that the optimized teaching method has a positive effect on the transformation of students’ learning attitude towards the C programming course.

Figure 9.

Question 6 statistics

Figure 10 shows the results of “I am satisfied with the current classroom model based on the optimization of teaching method”. More than 85% of students are satisfied with this teaching model.

Figure 10.

Question 7 statistics

Figure 11 shows the statistical results of “I am willing to accept the extension of this teaching mode to the teaching process of other courses”, 94% of students strongly agree or agree with the addition of AI-assisted optimization teaching methods in other courses, indicating that the optimization teaching method has a certain positive effect and attraction on students’ learning, and students have a high acceptance of this AI-assisted teaching method.

Figure 11.

Question 8 statistics

Therefore, it can be concluded that the implementation of the Optimisation Teaching Method is effective in the C Programming course in the Computer Science class of a vocational school. After the implementation of the optimisation teaching method, the student’s interest in learning and their participation in the classroom has increased to some extent, and this AI-assisted teaching method has received more positive feedback, showing the effectiveness of the method in secondary schools. Therefore, based on the positive feedback after the implementation of the optimised teaching method, it can be concluded that the implementation of the optimised teaching method in secondary schools is effective, and the method will continue to be explored and optimised in the future in order to further improve the learning outcomes and vocational competence of students.