Research on the Teaching Reform Path of Colleges and Universities in Transportation and Water Transportation Education in the Era of Artificial Intelligence

Professional positioning and training objectives must be consistent with the current standard of the transport industry’s demand for talent. At present, in the economic globalisation, integration, multimodal transport, and logistics, under the booming trend of international multimodal transport, mainly maritime transport, the rapid development of China’s transport industry ushered in the era of great development can organise and manage a variety of transport modes of transport professional management personnel demand surges. While colleges and universities specializing in maritime transport typically focus on shipping, students generally lack interest in other modes of transportation and a deeper understanding. It is suggested that the transport majors of maritime colleges and universities should be positioned as transport majors with distinctive maritime characteristics, aiming at cultivating future multimodal transport leaders with maritime transport as the mainstay. The teaching of basic knowledge about other modes of transportation should be strengthened to fill the gap in demand.

First of all, to do three balances: traffic and water transport education and transport professional management, but from a practical point of view, one also needs to have a strong engineering background, so in the selection of public foundation courses and professional foundation courses to pay attention to the engineering knowledge and management knowledge of the balance; shipping management students graduated from most of the first directly engaged in the business operations, to accumulate some practical experience, there will be a better development, but from the undergraduate level of training objectives, the need to train them into the future industry or enterprise managers and leaders, so in the curriculum to pay attention to the business operations and professional theory of the balance. However, from the point of view of the cultivation goal of undergraduate level, it is necessary to cultivate them to become the managers and leaders of the future industry or enterprise, so pay attention to the balance between business operation and professional theory in the arrangement of the curriculum.

Because of the late establishment of transportation and water transport education, colleges and universities do not have an advantage in terms of teachers’ strength, which can be remedied by increasing the training of teachers and rationally organizing the teaching process. For professional courses in marine colleges and universities, transportation is also a key factor in the development of professionals. With the support of the school, take the “bring in, go out” way to give teachers the opportunity to the development of a long history, professional development of mature, well-known institutions to learn from the experience to the enterprise to work and practice, to improve the level of professional teachers. Encourage and direct students to participate in more off-campus internships to enhance their knowledge of the specialty. Introducing outstanding professors and transportation experts from well-known institutions as visiting professors of our school to give regular reports and lectures to students so as to increase teachers’ knowledge and stimulate students’ interest in their specialties.

Based on the above actual research and analysis of the online Q&A function of the teaching website of a provincial university’s high-quality courses, the intelligent interactive platform of the teaching website of the courses based on the course knowledge base is constructed with intelligence, and the platform is based on the B/S mode, and the network structure is shown in Fig. 1.

Figure 1.

Platform network structure

The platform allows students to ask questions in the form of natural language (Chinese or English, etc.) interrogations and returns relevant answers to students in a timely manner. If a relevant answer is not found, the question is sent to the course instructor’s email address or posted on message boards and forums for an answer and reply. Through the development of such a platform, the aim is to reform the online Q&A mode of university course websites, especially the boutique course websites, to fully mobilise the autonomy of students’ learning and to improve the resource utilisation rate of the course websites themselves.

The platform is based on course knowledge base search, supplemented by email, message board forums, and other means of answering questions. The core functions of the platform are as follows: 1)

Accurately understand the students’ questions.

The development of an online intelligent interactive platform for the course involves analyzing and comprehending the questions posed by students, as well as retrieving the course’s knowledge base. Among them, the key technologies involved in the analysis and understanding of the questions include text segmentation and keyword extraction and expansion, combined with general dictionaries, professional dictionaries and semantic dictionaries (using the Expanded Synonym Thesaurus provided by the Information Retrieval Research Office of the Harbin Institute of Technology), in order to achieve the analysis and understanding of the students’ questions.

The key technology involving the retrieval of the course knowledge base is mainly the calculation of sentence similarity, which is achieved by improving the general keyword method and the vector space model method (introducing positional information and conceptual semantic information, respectively) and considering the keyword method and the vector space model method according to a certain combination of weights to achieve the similarity calculation of the student’s questions and the questions in the course knowledge base. In terms of implementing the self-learning function of the platform, the machine learning approach is used.

The platform consists of three main parts: problem understanding, document retrieval, and answer extraction. The structure of the system is shown in Figure 2, which also reflects the platform’s business processing flow.

Figure 2.

Platform architecture

The implementation of the platform makes full use of existing Internet information retrieval tools (search engines such as Google and Baidu), open-source machine learning software packages (learning models such as SVM and CRF), and commonly used text clustering algorithms and concept lattice building algorithms. In terms of the required basic language technology processing modules (segmentation, lexical annotation, named entity recognition, dependency analysis, etc.) and related corpus resources (Chinese question sets, etc.), this is achieved by calling the segmentation component of the Institute of Computing of the Chinese Academy of Sciences (ICCAS) and the related modules of the LTP platform of the Information Retrieval Research Centre of the Harvard University of Technology (HUT).

Semi-supervised algorithms have greater advantages in specific data retrieval, mainly due to the fact that the algorithm can be more intuitive and easy to complete the information matching in the process of information retrieval, with strong interpretability. Therefore, it has been applied in many different types of knowledge retrieval with more satisfactory results [19].

In the process of knowledge retrieval based on natural language processing, the entire data is constructed as an undirected graph using a semi-supervised algorithm. In the graph that was created, every point represents a data point, and the connecting line between two data points is utilized to describe the similarity between the data points. Setting X describes the weight matrix of the connecting line in the graph, and line x is used to describe the weight of the connecting line between data point j and data point k. Assuming that there is no connecting line between the two data points mentioned above, x_jk = 0, Setting L is a neighbouring data point. Assuming that data point k is in the neighbourhood of data point j, x_jk = 1, otherwise x_jk = 0. Assuming that the above relationship is described using Gaussian kernel, the following conclusions can be obtained: (1)xjk=exp(−‖yj−yk‖2ζ2)

In the process elaborated above, the semi-supervised learning algorithm is introduced in the streaming condition. The so-called flow shape condition means that in the case of data samples with relatively high data complexity, they are stored in samples with relatively low data complexity. Therefore, if the above stream shape assumption holds, the knowledge retrieval method based on natural language processing can be implemented in the sample space with relatively low data complexity. Knowledge retrieval based on natural language processing using the method elaborated above can avoid the limitations of knowledge retrieval brought about by excessive data complexity, which fully demonstrates the superiority of the decision function in semi-supervised algorithms [20].

As a result, the data samples in the neighborhood region have a strong similarity. The computational complexity of the knowledge retrieval method based on natural language processing is also relatively high, and it is necessary to establish an undirected graph corresponding to the data points for the above knowledge retrieval method, obtain the Laplace matrix of the corresponding data through the operation, and get the data related to the Laplace characteristics according to the obtained results, assuming that the undirected graph of the above data has a high degree of completeness.

In the process of constructing an undirected graph based on the data samples, the similarity of the two data points can be customised, and the Gaussian kernel can be set to define the similarity, and for the process of label transfer between the data samples, the label transfer probability matrix can be established, and the size of the corresponding matrix is set to be (M+V)×(M+V) , so that the following results can be obtained: (2)Ujk=Q(j→k)=xjk∑l=1M+V

The specific steps for the undirected graph using the above method are shown below.

Input: a set E consisting of all data samples, including a set M consisting of labelled data and a set V consisting of unlabelled data, and a set D consisting of data attributes of different categories.

Output: labels of the data samples that are not labelled.

The algorithm is as follows: 1)

Construct an undirected graph H=〈W,F〉 of the data samples according to the customised way of calculating the similarity between the data samples, and obtain the corresponding similarity matrix X.

Assuming that the time complexity takes more than P(lp2) in the above process of knowledge retrieval based on natural language processing, since the mark propagation method belongs to the knowledge retrieval method in the direct push mode, it is necessary to test the algorithm again in the process of knowledge retrieval each time, so it is necessary to extract the subject words in the data samples in a short period of time to reduce the computational complexity in the process of knowledge retrieval.

Combining the characteristics of each domain knowledge, the user’s question is firstly matched with the question in the expert Q&A database to find out whether there is the same similar answer. If the matching fails in the expert Q&A database, then according to the type of question, the answer is retrieved in the enterprise knowledge base, or the design principle is deduced as the answer for the user by using the TRIZ contradiction matrix, respectively. The answer retrieval framework is shown in Figure 3:

Figure 3.

Answer retrieve architecture graphic

After the interrogative processing, the user’s interrogative query is submitted to the answer retrieval module, and the processing flow of the answer retrieval module is as follows:

According to the question information extracted by the question analysis module, build a structured query statement.

① Similarity matching is performed in the expert Q&A library to find out whether there is an answer that meets the conditions. If the search is successful, turn to ⑦. Otherwise, turn to ③ processing.

③ Judge the question classification. If the question category is the concept definition category, turn to ④ processing; if the question category is the entity data category, turn to ⑤ processing; if the question category is the design principle category, turn to ⑥ processing.

④ Query the answer to the question in the concept database through rule matching. Go to ⑦.

⑤ Query the question answer in the design knowledge base by rule matching. Turn to ⑦.

⑥ Use the TRIZ contradiction matrix to interact with the user to reason and find the design principle type answer.

⑦ Return the user’s answer.

Through the interrogative processing, the interrogative sentences proposed by the user are subjected to disambiguation, lexical annotation, dependency analysis, extraction of interrogative features and classification of interrogative sentences, and this interrogative sentence information is the basis of the answer retrieval module. If the keywords in the question sentences are directly used for matching, the semantic information of the question sentences will be lost, and at the same time, the domain knowledge in the enterprise has a certain structure, so establishing a certain structure for the information of the question sentences and retaining their semantic information will help the efficiency and accuracy of the answer retrieval. According to the information that can be extracted from the interrogative sentence, we establish the query information class, which mainly includes conceptual entities, extended concepts, question categories, dependency pairs, question words, interrogative cores, interrogative verbs, and interrogative keywords.

In the calculation of the similarity of interrogative sentences in this paper, we believe that all of the following factors should be given consideration: 1)

Entity of the question. The entity asked by the user is the most important factor of the interrogative sentence, and if the interrogative entities are not the same, then the similarity of the other parts of the interrogative sentence should be disregarded.

According to the above three factors, this paper proposes an improved algorithm that combines the semantics of question keywords and dependency relations.

The questioning entity determines the centre of attention of the questioning, and in the query information class, we define two types of vocabularies, conceptual entities and extended concepts. For the user question S₁ and the expert question and answer bank S₂, there are concept entity sets C₁, C₂ and extended concept sets E₁, E₂ respectively: (4)C1={c1,c2,…,cm} E1={e1,e2,…,ej} C2={d1,d2,…,dn} E2={f1,f2,…,fk}

The lexical similarity function Sim_w is used to calculate the proportion of identical words in two word sets: (5)Simw(C1,C2)=SameWord(C1,C2)Max(m,n)

Where Same Word (C1,C2) means C₁, C₂ contains the same number of words.

Define the conceptual similarity Sim_c formula as follows: (6)Simc(S1,S2)=αSimw(C1,C2)+∅Simw(E1,E2)

where α = 0.7, β = 0.3 are used to distinguish the weights of entity concepts and extended concepts.

The question keywords in the query information category are the other words of the question after removing the stop words and conceptual entities, and the semantic relevance of these words also affects the similarity of the whole sentence. For two words W₁, W₂, W₁ have n terms: S₁₁, S₁₂, …, S_1n; W₂ has m terms: S₂₁, S₂₂, …, S_2m, then W₁, W₂ semantic similarity is calculated as: (7)Simw(W1,W2)=maxi=1,n,j=1,mSim(S1i,S2j)

Where the Sim(S1,S2) -sense item similarity is calculated as: (8)Sim(S1,S2)=∑i=14βi∏j=1iSimj(S1,S2)

Where β₁ = 0.5, β₂ = 0.2, β₃ = 0.17, β₄ = 0.13, Sim₁ is the first basic semantic similarity, Sim₂ is the other basic semantic similarity, Sim₃ is the relational semantic similarity, and Sim₄ is the relational symbol similarity.

Sentences S₁, S₂, S₁ have keywords {W11,W12,…,W1m} , S₂ have keywords {W21,W22,…,W2n} , and using the results of lexical question similarity calculation, we define the semantic similarity formula for sentences S₁, S₂ as follows: (9)Sims=∑i=1mmaxj=1..n(Simw(W1i,W2j))m

Much of the sentence structure is reflected in the dependencies of the sentences, and the similarity calculation of the dependencies of two interrogative sentences is also incorporated into the interrogative similarity. Among all pairs of dependencies, we selected two categories as valid dependencies: those that contain domain vocabulary from the domain lexicon, and those that contain the core HED of the interrogative sentence. Both types of dependencies are closely related to the semantics of the question. For dependency pairs D1(W1,W2) and D2(W3,W4) the similarity is calculated as: (10)Simp= max(Simw(W1,W3),Simw(W1,W4)) +max(Simw(W2,W3),Simw(W2,W4))2

Then interrogative sentences S₁, S₂, S₁ have valid dependency pairs {D11,D12,…,D1m} , S₂ have valid dependency pair {D21,D22,…,D2n} , and the dependency similarity formula for interrogative sentences S₁, S₂ is defined as follows: (11)Simd=∑i=1mmaxj=1...n(Simp(D1i,D2j))m

By calculating the similarity of each element of the interrogative sentence above, we set the formula for calculating the similarity Similarity(S1,S2) of the interrogative sentence for two interrogative sentences S₁, S₂ as follows: (12)Similarity(S1,S2)=w1Simc(S1,S2) +w2Sims(S1,S2)+w3Simd(S1,S2)

Where w₁ = 0.5, w₂ = 0.2, w₃ = 0.3. In order to prioritise accuracy, the similarity threshold is set to 0.85. i.e., when similarity is higher than this threshold, we consider that the user’s question matches with the question in the expert’s Q&A database, and return the answer from the Q&A database to the user.

Queries that cannot be matched to the same question through the Expert Q&A database are submitted to the Knowledge Base or TRIZ for answer extraction based on the question category. Questions with the question category of conceptual entities are searched for answers in the conceptual library, and those with the category of entity data are searched in the design knowledge base. Since the knowledge base is stored using a relational database, the retrieval of answers to the knowledge base is mainly a structured query using the information processed on the question and transformed into the database SQL language of the knowledge base by rules.

In this paper, 6 groups of questions are randomly selected from inside the question bank by random sampling. Each group has 600 questions, a total of 600 questions, and these 600 questions are used as the experimental data source of this experiment. The experiment is divided into 3 experiments to compare and analyse the effectiveness of knowledge retrieval methods and optimisation by comparing the relevant indicators before and after optimisation. The first experiment is to use the basic attribute template and direct component. The second experiment is to optimize subject-object reversal, and the third experiment is to optimize alternative keywords. The article refers to some commonly used evaluation indexes in other experiments and then combined with the characteristics of the retrieval mechanism of this paper, a total of four evaluation indexes are given, three basic indexes: the missed detection rate (T₁), the correct rate (T₂), the quasi-checking rate (T₃), and the comprehensive indicator (T₄): it is used to take into account the correct rate and the accurate detection rate, and indicates the quality of the retrieval method.

1)

Test experiment results

In this paper, the relevant indicators of six experiments were recorded separately, and their average values were calculated. Through the comparative analysis of the relevant indexes of each experiment, the retrieval mechanism and optimisation method are analysed and summarised.

Experiment 3: The third experiment is to add alternative keyword optimisation in the political standard library to test the retrieval mechanism accordingly, and the experimental test records are shown in Table 3. As can be seen from the table, the correct rate and comprehensive indexes are improved compared to Table 1, respectively, 0.038 and 0.036, the omission rate has decreased, and the relevant indexes are improved a little higher than the experiment two improvement. It shows that adding alternative keywords has some effect, and the effect is better than subject-object reversal optimisation, which solves the situation that the subject in some interrogative sentences has the same thing and a different name as the subject in the library.

2)

Comprehensive comparative analysis

In this paper, a comprehensive comparative analysis of the average values of the evaluation indicators of this experimental design in the three experiments is carried out, and the comprehensive comparison of the average values of the indicators is shown in Figure 4. As can be seen from the figure, on the basis of Experiment 1, Experiment 2, and Experiment 3, the three optimization methods used alone, the correct rate T2 compared to Experiment 1 have been improved respectively by 0.033 and 0.038, the leakage rate T1 compared to Experiment 1 have been reduced, the quasi-checking rate T3 compared to Experiment 1, although in decline, but the comprehensive index T4 has been improved, proving that the optimization methods have a certain degree of feasibility, indicating that the three kinds of optimisation methods are feasible, indicating that the combination of the three optimisation methods has a significant effect.

Figure 4.

Three experiments were combined