Functional Analysis of English Carriers and Related Resources of Cultural Communication in Internet Media

3.1 Explicit and Invisible Strategies of English Culture Teaching in the Network Platform

Since the reform and opening up, China’s English education has been in the process of development. It has made efforts to meet international standards. However, students and teachers pay less attention to communicative teaching in educational activities. In addition, students are also in a social environment where English is not often used. The influence of these factors has led to the low oral English ability of Chinese students (Kumayas & Lengkoan, 2023). Besides, the cultural knowledge in the current English textbooks of all stages in China is dominated by English-speaking countries, ignoring the communicative role of cross-cultural teaching. This leads to the insufficient ability of Chinese students to use English to express their local environment and culture. Therefore, students should be aware of the differences between Chinese and Western cultures and have a more thorough understanding when reading first-hand English materials. Mastering communication etiquette and taboos in life and work can effectively improve the ability of cross-cultural communication. The cultural aspect of English teaching usually includes the basic theories and knowledge of the history, politics, economy, diplomacy, social culture, literature, and other aspects of British and American countries. In cultural teaching, students can better accept the skills and knowledge of English listening, speaking, reading, writing, and translating. Combining technical knowledge and cultural content, the ultimate goal of English teaching is to cultivate high-quality, high-level students with the ability to engage in translation, research, teaching, and management work. The contents of English skills application teaching and cultural knowledge teaching are shown in Figure 1.

Figure 1

The content of English skill application teaching and cultural knowledge teaching.

As shown in Figure 1, students should be able to understand some Chinese and foreign cultures and ideas in teaching and improve their humanistic, artistic, and aesthetic qualities. To cultivate and improve the cultural quality of English majors, students should establish correct social, historical, and life values. From the dominant perspective, first of all, English education shows a trend of increasing knowledge application skills in the distribution of courses, while cultural and literature courses decrease. Second, English skill learning courses give too much attention to grammatical structure analysis, which makes the content difficult and reduces the interaction between teachers and students. This also led to an inactive learning atmosphere in the classroom. Finally, there is a lack of hierarchy in English teaching. Students have insufficient time to sort out and understand the teaching content. If culture is divided into utensil culture, institutional culture, and conceptual culture, the English culture teaching in China mostly stays on utensil and institutional culture. Its teaching rarely involves the content of the third stage. From the perspective of invisibility, first, it is difficult to give full play to students’ advantages of online autonomous learning. Second, student-oriented English culture teaching activities show the characteristics of shallow content, simple form, and passive acceptance by students. Finally, students in autonomous learning have few opportunities to practice. The difficulties faced by English culture teaching at the present stage are shown in Figure 2.

Figure 2

Difficulties in English culture teaching at the present stage.

With the rapid development and popularization of the Internet in the new century, podcasts, blogs, Massive Open Online Courses (MOOC), etc., provide new channels for digital and mobile English learning. In educational text data, there are a large number of academic terms that can reflect the current research content and theme of the text. These terms can usually be used as keywords to summarize an article. The keyword extraction method for word graphs is based on word nodes, calculating word weights to generate keywords. Compared to package and topic models, extracting keywords from the perspective of words is better. Therefore, this article chooses the text ranking graph model as the basic model for keyword extraction. With the rise of pre-trained language models, many researchers in the field of keyword extraction have combined pre-trained language modeling with unsupervised keyword extraction models, which not only ensures the performance of keyword extraction but also enriches semantic information, to some extent solving the problem of low accuracy in traditional unsupervised keyword extraction models. This article proposes an improved text ranking keyword extraction model based on pre-trained language models, aiming to improve the accuracy of educational text keyword extraction and enhance the efficiency of extraction tasks.

This also brings new opportunities for English learning. Relying on the gradually developing network technology, teaching resources on the network show the characteristics of wide richness and high convenience, which play an important role in English teaching. Internet English teaching supplements the cultural content that students learn in the classroom and can ensure the complete transmission of teaching information. Second, there are many forms of network resources, such as text, voice, video, and image. The presentation of network resources is more vivid and diverse. Knowledge is more easily accepted by students in this way. Teachers can arrange the teaching contents in a more orderly way by relying on the network platform to improve the coherence and systematization of English teaching. Finally, the network platform can strengthen the interactivity of English teaching and activate the classroom atmosphere.

3.2 Application of the BERT Model in the Standards of English Online Teaching Resources

This method classifies the text of audio and video resources by designing an optimized BERT model and designs a Text Rank keyword extraction model that integrates external knowledge and semantic feature weights to extract the knowledge points contained in the audio and video resources. Finally, the classification label and keyword extraction result are used together as the label of the audio and video resources.

Despite the growing importance of network resources for English culture teaching, the organization of the teaching resource database is still in a low efficiency and slow query application mode. Knowledge atlas has become a hot spot to improve the structure of English teaching. Representing resource content information through tags is the basis of building a knowledge map. This research will use the method of deep learning to build a label-generation model of English online teaching resources. The process of generating resource annotations from this model is shown in Figure 3.

Figure 3

Model flow of tagging generation of English teaching network resources.

Figure 3 shows that the educational resource annotation model built in this study mainly relies on two tasks: keyword extraction and classification, which are completed by Text Rank and BERT models, respectively. BERT is a pre-training language model for two-way feature representation, which can simultaneously obtain the context feature representation of text. The core of BERT is the transformer structure. The bidirectional transformer encoder is composed of multiple layers. The characteristics of the basic transformer encoder enable BERT to better learn the context information in text data. The bidirectional network structure ensures that BERT can mine more semantic information. The structure of the transformer is shown in Figure 4.

Figure 4

Transformer structure.

As shown in Figure 4, the transformer encoder is composed of a multi-head attention mechanism (MHAM) and a fully connected feed-forward network (FCFFN). Each layer has a residual connection and normalization layer. Assuming that the input of each sub-layer is x, the output of the sub-layer is expressed as equation (1)

The MHAM model is composed of multiple scaled dot product attention (SDPA) models, and its mathematical expression is shown in equation (2)

In equation (2), the input of attention is queries ( Q ) with dimension d q , keys ( K ) with dimension d k , and values ( V ) with dimension d v . The mathematical expression of the MHAM model is equation (3)

First, a linear mapping is performed for the matrix with input dimension d mod el Q , K , V . The parameters of the linear transformation matrix are W i Q ∈ R d model × d q , W i K ∈ R d model × d k , and W i V ∈ R d model × d v . Q ∈ R d model × d q , K ∈ R d model × d k , and V ∈ R d model × d v are obtained after mapping. Then, the new matrix of SDPA is used to calculate, and the calculation results are obtained after performing the above steps for h times. Finally, the merged results are linearly transformed. The linear transformation matrix parameter is W O ∈ R h d v × d model . The output result after linear transformation is matrix M h d v × d model . After the MHAM calculation, matrix M obtains the semantic features of different spaces of the text and captures the context information of each word. FCFFN is composed of two layers of a linear fully connected network, and the linear transformation in the middle depends on the ReLU activation function, as in equation (4)

To apply the BERT model to video resource annotation, it is necessary to preprocess the teaching videos first. After converting the audio of the video into text data, this study used the Chardet tool to detect the text data format and unified the text data into UTF-8 encoding. After text preprocessing, text vectors are extracted from the BRT model. If the word vector matrix is used as input to the model, the relationship between the original text sentence and the vector matrix is expressed as follows (5):

where M represents the word vector matrix, S represents the original sentence, N represents the number of words in the sentence, and d represents the vector dimension of each word (Guo et al., 2020).

3.3 Research on the Model of Educational Resources Labeling in English Online Media

To achieve the key word extraction of the tagging model of English online teaching resources, the Text Rank model used in this study is essentially a weighted undirected graph model. It is assumed that any sentence in the sentence set T = { S 1 , S 2 , … , S n } is expressed as S i = { V 1 , V 2 , … , V m } . V represents the words in the sentence. Therefore, text can be represented by a graph model G = ( V , E ) . V refers to the set of all word nodes in the word graph model. E refers to the weight set of edges between each word node in the word graph. The Text Rank value of the word node V i is expressed as equation (6) (Jiao et al., 2023)

In equation (6), S ( V i ) is the importance score of the word node V i . d is the damping coefficient. In ( V i ) and Out ( V i ) represent the degree of entry and exit of word nodes in the word map. As a recursive formula model, the initial value of each word node in the original Text Rank model is the same. In this article, the research object in keyword extraction is educational text data. Educational texts contain subject-specific nouns, which are less likely to appear in ordinary texts. However, they may appear repeatedly in educational texts, and these proper nouns can reflect the main content of educational texts. If the unified initial weight is used to reconstruct the word map when performing the keyword extraction task of educational text, it will inevitably affect the importance ranking of proper nouns and thus affect the whole keyword extraction result. Therefore, a method for calculating the initial weight of multi-domain words based on an external proper noun database is proposed. This method improves the accuracy of the model by accurately assigning the initial weight of the word map. The optimized version of equation (6) is shown in equation (7)

In equation (7), λ represents the thesaurus selection parameter. S com ( V i ) refers to the weight of the current word node in the general thesaurus. S dom ( V i ) is the weight of the current word node in the proper noun database. N represents the number of word nodes in the word map. M 1 and M 2 are the number of words appearing in the general vocabulary and special vocabulary in the current word map. The Text Rank model is essentially calculated on the basis of a word graph. Although this model increases the importance of proper nouns through initialization weight, the importance of proper nouns is too varying. Therefore, the research improves the probability transfer matrix of Text Rank through the BERT pre-training model and distinguishes the importance of words more accurately by integrating the meaning of words. First, the degree of semantic correlation between word vectors is calculated by the cosine similarity formula, and equation (8) is obtained as follows:

In equation (8), V i , V j represent two word nodes. V ¯ i and V ¯ j are two-word vector expressions. In the weight allocation, the greater the semantic difference between two adjacent word nodes, the higher the transfer probability should be given. Therefore, the probability calculation formula of the transition from word node V i to V j is expressed as P sim ( V i , V j ) = 1 − sim ( V i , V j ) . Then, the external knowledge weight and semantic weight are combined to comprehensively calculate the jump probability of nodes, as in equation (9)

The formula after improving the weight transfer matrix is equation (10):

After the probability transfer matrix is obtained, the Text Rank value of the word node can be finally calculated, as shown in equation (11):

Finally, when the research and construction of the English teaching network resource labeling method is applied to the online teaching platform, the relationship between administrators, teachers, and students and the system functions are shown in Figure 5.

Figure 5

Application scenario of resource annotation model in online teaching platform.

In Figure 5, the model built by the research takes into account the actual needs of teaching tasks. The design of educational video resource marking systems can provide users with convenient educational video annotation, result display, data storage, and other services. Thus, various personalized online teaching platforms can make better use of educational video data. The educational video resource tagging model includes module design and function construction, including data preprocessing, text classification, keyword extraction, and result display functions.

4.1 BERT Classification Model Experiment Based on Network Education Resources

To test the constructed annotation model of English teaching resources, the research took all English course teaching videos in the online teaching resources platform of a university as raw data. The audio in the video was extracted and stored separately. The experiment used the format factory audio and video conversion tool to save the audio track as a wav file. Intelligent speech recognition service was used to convert teachers’ teaching content into text data. The model parameters of the experiment are shown in Table 1.

In the experiment, the structure of the BERT classification model was composed of a 12-layer bidirectional transformer decoder. Its hidden layer dimension was 768 with 12 attention mechanisms, and the overall parameter quantity was about 110 M. The maximum length of the text sentence was 510 words, and the learning rate was 5 × 10⁻⁵. The number of samples for each batch during training was 64. To verify the effectiveness of the BERT classification model in English teaching network resources, that experiment first sets up the classification performance of the model in different teaching text interceptions. Since the length of text in the teaching video resources was generally more than 510 words, the study adopted three methods to obtain text data of different lengths. They were the first segment interception method, the first and last combination interception method, and the direct interception method. The macro precision of the classification model under different interception methods is shown in Table 2.

Macro-P, Macro-R, and Macro-F1 in the table represented the precision rate, recall rate, and F1 value of performance indicators. The data in Table 2 showed that under the same model, different data interception methods directly affected the final effect of the model. From the two results of the direct interception method, both the beginning and the end of the document contained information that could summarize the document. The interception method that combined the first three sentences and the last two sentences performed better, reaching a macro-F1 value of 91.3%. Meanwhile, its precision and recall rates were 91.4 and 91.1%, respectively, which were also higher than other data interceptions. In summary, the article believed that in the text data of English education in a university, most of the course content was contained in the teacher’s opening remarks. At the end of the course, the teacher would also review and summarize this lesson. The course content could be highly summarized by combining the first three sentences and the last two sentences. The number of text characters obtained varied from the original article, ranging from 200 to 350 words. After the experiment of text interception strategy, this study compared the performance of different classification models.

Figure 6 shows that the classification results of the Fast Text model were only second to those of Text convolutional neural network (CNN) and BERT series models. Its F1 value was 87.8%. The F1 value of Text CNN was only 0.5% different from the original BERT model. The long short-term memory (LSTM) model had the worst performance. The F1 value of Binary LSTM was 1.4% higher than the original LSTM model but still 4.1% lower than the original BERT model. The BERT-T model proposed in the article had the best effect in the experiment, with the macro-F1 reaching 91.2%. It was 0.8% higher than the macro-F1 of the original BERT model and 0.5% higher than the macro-F1 of the Ro BERTA model. The reason was that the CNN structure had excellent ability in local information extraction, which could more effectively extract the information in the BERT output word vector. In order to highlight the accuracy of the BERT-Text Rank model in video text classification and recognition, 50 data samples were randomly selected from the university course text data set to compare the classification accuracy of traditional classification methods and the BERT-Text Rank model studied, as shown in Figure 7.

Figure 6

Performance comparison of different classification models.

Figure 7

Comparison of classification accuracy between traditional clustering algorithm and BERT-Text Rank model studied.

The traditional clustering algorithm in Figure 7 was a single K-means algorithm, with computer, physics, history, and others numbered 1–4. From Figure 7, it can be seen that out of 50 test samples, the traditional method made 16 judgments with a recognition accuracy of only 68%, while the BERT-Text Rank model made only two judgments with an accuracy of 96%. The data showed that the BERT-Text Rank model had higher accuracy in determining the category of video text samples.

4.2 B-Text Rank Keyword Extraction Application Experiment of Online Education Resources

This research took the Wiley InterScience (WIS) database and the English literature data of a university as the experimental objects and validated the semantic weighted keyword extraction model. The WIS database contained social science and natural science journals in English. This experiment first compared the B-Text Rank keyword extraction model built in the study with term frequency-inverse document frequency (TF-IDF), Yet Another Keyword Extractor (YAKE), the original Text Rank, and EmbedRank models. Table 3 shows the comparison of macro precision values of 3, 5, and 7 keywords extracted from two data sets by different models.

The comparison of macro recall rate values in Table 3 showed that the more keywords extracted, the greater the value of each model. In the two datasets, the keyword extraction method with external knowledge weight proposed in this article was superior to other baseline models in obtaining the macro-R value. When the number of keywords extracted was 7, the B-Text Rank model proposed in this article was 0.47% higher than the original Text Rank model. From the comparison results of macro precision in the table, it could be found that the keyword extraction method proposed in this article with external knowledge weights on the two datasets had macro P values of 25.61, 21.77, and 18.44% in the TOP3, TOP5, and TOP7 cases, respectively, which were better than other baseline models. When the number of keyword extractions was set to 3, the Text Rank model with external knowledge weight proposed was 0.84% higher than the original model. Finally, the F1 value of different algorithms is shown in Figure 8.

Figure 8

F1 value of different algorithms. (a) College English literature database and (b) Wiley science database.

The two datasets in Figure 8 showed that the keyword extraction method with external knowledge weights was better than other baseline models. When five keywords were selected, the B-Text Rank model was 0.56 and 0.69% higher than the TF-IDF and YAKE based on statistical methods and 0.19% higher than the original Text Rank model. This group of experiments set up two control groups to verify the feasibility and efficiency of the interception method proposed in keyword extraction. The setting purpose of control group 1 was to verify the feasibility and effectiveness of the text rank model integrating semantic weight. Control group 2 was to verify the impact of data interception methods on the keyword extraction process. The specific results are shown in Table 4.

In Table 4, the performance of the B-Text Rank was improved compared with other baseline models in terms of macro-F1 value indicators. In the experiment of control group 1, the B-Text Rank was 2.19 and 2.85% higher than the traditional Text Rank in the two models. This showed that the fusion of semantic weights was feasible and improved the performance of the model. In control group experiment 2, this study showed the best performance in extracting three keywords, with macro-F1 values of 21.81 and 22.6% in the two datasets. In conclusion, B-Text Rank performed well under the comprehensive consideration of model extraction performance and efficiency. At last, the experiment applied the B-Text Rank resource annotation model proposed in the study to the MOOC forum management of the university. It also analyzed satisfaction evaluation from the feedback of the teacher administrator. Details are shown in Figure 9.

Figure 9

MOOC administrator’s satisfaction evaluation on traditional classification methods and optimized management methods.

The new MOOC intelligent management labeling method fed back its satisfaction evaluation through the administrator’s scoring behavior. The average satisfaction of teachers’ overall service was 84.95%, higher than 5.68% in the old system. It meant that the staff generally recognized the effect of model classification on the classification of comments in MOOC forums. In terms of teaching effect, the satisfaction of teachers in the new system was 82.64 ± 1.86%, while that in the old system was 80.41 ± 0.79%. In terms of teaching feedback and communication effect, teachers’ average satisfaction increased by 11.6 and 7.9%. In terms of fault resolution, the satisfaction of teachers in the new system was 84.95 ± 1.34%, and that in the old system was 79.27 ± 1.83%.