Intelligent Recommendation System for English Vocabulary Learning – Based on Crowdsensing

Adaptive Learning originated in the 1970s, studies and simulates people's learning from the perspective of information processing and is essentially a computer program [7]. This program can generate new rules in the process of problem solving, and add them to the program, so as to solve similar problems more effectively. It mainly includes the following four aspects:

Change what: Adaptive learning can change the learning contents by distinguishing the difficulty level of a task, or change the form of representation and path selection of the content [8].

As shown in Figure 1, the four-dimensional view of adaptive learning very intuitively shows a core problem in the design of adaptive learning system. It can be divided into learner-oriented and educator-oriented adaptive learning. The latter of these is characterised by a big conflict between curriculum management and the existing teaching system, and there are few mature products at present. As for the former, the system is familiar with the functions of acquiring interactive data of learners, creating a learning model, selecting information and displaying information. Learner-oriented adaptive learning can be realised through the cycle of four basic stages: acquisition, analysis, selection and presentation.

Fig. 1

Mobile Crowd Sensing (MCS) [9] refers to collecting data by using smart mobile device sensors in physical space and mobile social networks in cyberspace. Owing to its inherent high mobility and scalability, it has been widely used in many sensing applications [10], as shown in Figure 2.

Fig. 2

Group intelligence has the following two characteristics [11, 12]: First, there is no central control, but it is distributed in such a manner as to be suitable for the working state under the network environment, and will not affect the solution of the whole problem due to the mistakes of decision-making and data anomalies. Second, groups are self-organised. Each individual in the group can communicate indirectly for information transmission and cooperation, and thus each individual can change the environment. However, the complex behaviour of the group is reflected by the intelligence in the process of individual interaction. This kind of wisdom emerging from the group is beyond the wisdom of its constituent individuals which is called group intelligence [13].

Group intelligence is applied in the educational environment, that is, individual learners have certain rules about their abilities and behaviours to learn specific knowledge, and they can communicate indirectly. Therefore, for new learners, it has great reference and application value. In the teaching process, the conclusion of group intelligence can be used to improve teaching methods, update teaching contents and share learning strategies, so as to help learners optimise the learning process and improve teaching efficiency.

The role of the system is divided into learner and manager. The main functions of learners include: personal information maintenance, uploading corpus, completing tests, and vocabulary learning. The main functions of the administrator are to review corpus, update thesaurus, publish tests, recommend new words and so on.

As shown in Figure 3, the architecture of intelligent vocabulary recommendation system is divided into three layers: user layer, business layer and data layer. The layer provides data storage services and undertakes the responsibility of ensuring the reliability and security of data. According to the actual demand of the system, data layer stores user information database, user log behaviour database, vocabulary database, corpus, comment data published by users, test data and so on. The business layer implements the core business logic of the recommendation system, including similar words’ mining and adopting a user-based collaborative filtering algorithm to recommend words to learners, where the clustering algorithm is used to locate the learning style of users and adjust the push mode. The user layer is responsible for the interaction between learners and the system, the server response to the user's request and the display of the results, such as word learning, registration, corpus uploading, testing and other functions. Moreover, all data of user behaviour generated by this layer will be recorded in the log database.

Fig. 3

At present, the open corpus data set is very rich, for example, Wikipedia has opened the download channel of its multilingual webpage data. Take news corpus as an example; there is a plethora of available data sets, including: English News Data of Reuters Corpora [24] (Reuters English news text and corresponding data of news category), data set of New York Times (registered developer account application api, which can be traced back to 1851, any searchable and indexable article), data set of 20Newsgroups [25] (about 20,000 newsgroup documents are collected, which are evenly divided into newsgroup collections of 20 different topics), etc. The experimental data set used in this paper is 3,000 news reports from the figure-eight website [26, 27].

In the process of processing the original corpus, there are many meaningless symbols and function words, and thus it is necessary to remove stop words, extract stem, restore word form and carry out other operations. Firstly, the Word_tokenize () function in NLTK package is used to segment the text data, and the information is processed into the smallest unit that the computer can handle. Similarly, NLTK package is adopted to remove function words (conjunctions, prepositions, articles, etc.) and punctuation marks from the corpus. First, by defining an empty list and traversing the segmented text list, the empty list is appended if the word does not exist in the deactivated word list. Secondly, Porter's stem extraction tool is used to extract the same stem from different inflections of English words. Finally, the WordNetLemmatizer function is used to restore the morphology. Thus, it is enough simply to understand the data structure of the incoming function. When these steps are implemented in Python language, importing packages and calling functions can be easily implemented. The example code is as follows:

fromnltk.corpusimportstopwords

In order to solve the “cold-start” of the system, the method of mining co-occurrence words and similar words of words is also adopted to realise mixed recommendation of words. The experimental corpus is trained by word2vec model of Gensim toolkit in python [28, 29], where the trained model is used to output similar words. It is very convenient to train word vectors in Gensim. We save the trained model and load it, and then call the most_similar () method to get similar words.

Taking the professional category as the boundary, the system pre-processes the corpus materials uploaded by administrators and users, and the number of core words obtained is limited (assuming that there are N words). By creating a co-occurrence word matrix, the common frequency of all words can be obtained. The higher the frequency, the greater the correlation between words, and it is classified as a list of preferred words, thus realising mixed recommendation. The specific implementation logic is shown in Table 1. The N*N matrix is constructed with all N core words as the number of rows and columns. The data on the diagonal line are uniformly defined as 0, and the data on both sides of the diagonal line are symmetrically distributed. The number in ath row and bth column is 10, and the number of common occurrences of words a and b across all the corpora is 10.

Owing to the large number of words, it is an important step to filter out unimportant words in the text and then calculate their total number. The text feature extraction algorithm TF-IDF is used to extract feature word vectors. The purpose of this algorithm can be understood as calculation of the importance of a word in a document [30, 31]. For the word i of a document j, the calculation of its TDIDF value is shown in Eq. (1): (1) w(i,j)=tf×idf=Count(i,j)Size(j)×log(NDocs(i,D)) w(i,j) = tf \times idf = {{Count({\rm{i}},{\rm{j}})} \over {Size({\rm{j}})}} \times \log \left( {{N \over {Docs(i,D)}}} \right) Term frequency (TF) calculates the frequency of word i appearing in document j. Inverse document frequency (IDF) is obtained by dividing the total number of documents N by the number of documents containing the word i, and then taking the logarithm of the quotient. We multiply TF and IDF, and the greater the TF-IDF value, the more important the word is to the article. Finally, the words with value of TF-IDF greater than 0.01 are extracted to form the feature word set of the corpus, which is realised using the following algorithm:

Vector = tfidfvectolizer (stop _ words = stopwords _ 1ist) # Call the tfidfvectolizer method to convert the input corpus into a word vector.

To build a word network, first, a common empty matrix needs to be built; we fill the first row and the first column of the matrix with feature words, and then calculate the common comments of feature words in each corpus. The main calculation of this segment code focuses on calculating the contribution times of feature words, which requires constant recursive operation. In order to improve the operation efficiency, the location of each feature word is stored in a dictionary {’feature words ’:[1,3,5,29,45,89]}. Then, we take out the corresponding position list of two feature words that need to be compared from the dictionary, turn the list into a set and use the method of intersection between two sets to ascertain the co-occurrence frequency of the two feature words. The final implementation algorithm is as follows:

defcount_matrix(matrix,formated_data):