Current Issue

The capabilities and use of online neural machine translation (NMT) systems have been rapidly increasing as evidenced by Naver Papago [1] and Google Translate [2]. In 2019,  Naver Papago reported over 10 million monthly active users [1]. In 2018, Google Translate reported translating “30 trillion sentences per year across 103 languages” [2].

An overwhelming majority of students utilize online translators (OTs) for language learning. Based on the results of large-scale questionnaires [3, 4, 5], 71% to 99% of students utilized OTs in their foreign language course. These students frequently used OTs to check words/phrases and perceived OTs as helpful for increasing target language vocabulary [3].

In this paper, we analyzed source and target text in a Korean-English NMT system to scaffold vocabulary learning. The source and target text was used to generate personalized English vocabulary lists based on the word frequency counts. These lists can help students to effectively study vocabulary words aligned with their academic goals. With these lists, students can strategically identify and study useful words needed to understand specialized texts and communicate technical ideas [6].

The rationale for our work is twofold. First, we understand that word selection is a crucial aspect of learning target language vocabulary. To select useful words, occurrence frequency must be considered as high-frequency words provide better returns for learning efforts [6]. Second, we noticed that users are unable to access word frequency statistics in widely used OTs. For Google Translate and Naver Papago, users can only select and save translated words/phrases in an online phrasebook.

To generate personalized word lists, we implemented a word counting module in our NMT system. First, we counted words in source and target sentences, respectively. After filtering out stop words [7] and high-frequency words [8], we generated two English vocabulary lists with the 100 most frequent words ranked from the highest to lowest frequency count. The first list was generated from 50 English articles with content related to the word ‘AI’ inputted into our NMT system. The second list was generated from 25 Korean articles with ‘AI’ content translated in our NMT system. When compared to other widely used academic word lists, there was low word overlap and these lists contained noticeably more area-specific words.

In this section, we provide a brief overview of our Korean-English NMT system, which consists of preprocessing, translation, and postprocessing.

The main contribution of this paper is generating personalized word lists in a translation system, aligned with the academic interests of individual OT users. To generate the word lists, we first extracted the words from the source and target text based on white space. We aimed to retrieve specialized words by filtering out 128 stop words [7] and 2,801 general high-frequency vocabulary word families [8]. The high-frequency word families are from the New General Service List (NGSL) developed for English language learners based on a “273 million word subsection of the 2 billion word Cambridge English Corpus” [8].

After generating lists based on the source and target text, we compared these lists to other widely used academic word lists. First, we examined the word overlap with the Academic Word List (AWL) [17] (composed of 570 word families), which is a general English for Academic Purposes (EAP) word list developed from a corpus of 3.5 million words from 28 subject areas across Arts, Commerce, Law, and Science. The NGSL and AWL are expected to cover 90% of the words in academic texts [18]. Second, we examined the word overlap with a science-specific word list (composed of 318 word families) developed from a corpus of academic texts from 14 science-related subject areas [19].

In our NMT system, users can access their word lists by pressing the ‘Statistics of Your Translation’ button to view the total word count and the 100 most frequent words ranked from the highest to lowest frequency with individual word counts as shown in Figure 1. Users can simultaneously view a list for their source text and a list for their target English text as shown in Figure 2.

Fig. 1. 
Translation Interface with Statistics Button in red.

Fig. 2. 
Word List Interface with source and target lists.

To generate a personalized word list for the source text as an example, we inputted 50 English articles with content related to AI from the most popular news websites in the United States. A total of 48,853 words from the 50 articles was inputted into our NMT system, and a list of the 100 most frequent words was retrieved as shown in Table 1. The list included AI-related abbreviations, nouns, adjectives, adverbs, and verbs. In the list, there were common AI terminologies (e.g., algorithm, neural) and trending words (e.g., amazon, pandemic, coronavirus). When compared to the other word lists, only 11 words overlapped with the AWL [17], and one word overlapped with the science-specific list [19].

To generate a personalized word list for the target text, we inputted 25 articles in Korean with content related to AI from popular news websites in Korea (e.g., The Kyunghyang Shinmun) and translated the text into English. A total of 18,333 words from the 25 articles were translated, and the statistics for the 100 most frequently translated words were retrieved as shown in Table 2. The list included AI-related abbreviations, nouns, adjectives, adverbs, and verbs. In the list, there were common AI terminologies (e.g., algorithm, robotics), trending words (e.g., pandemic, coronavirus), and context-specific words (e.g., hyundai, krw, k-pop). When compared to the other word lists, only 12 words overlapped with the AWL [17], and 4 words overlapped with the science-specific list [19]. When compared to the source word list, there was an overlap of 31 words with 2 words in a different form and 10 numbers.

The results from the experiment indicate that the personalized word lists generated by our NMT system include highly specific words directly aligned with the interests of individual OT users. These lists contain a variety of useful words such as terminology commonly used in specialized texts and words that provide insight into current area-specific trends. By referring to the lists, students can strategically identify words often used in their academic interest area.

To effectively study the words in the personalized lists, students can begin by reviewing the aspects (e.g., spelling, pronunciation, forms, definition), synonyms, and usage examples of each word using online dictionaries [6]. Then, students can practice retrieving word definitions with paper-based or digital word cards [20]. Finally, students can reinforce their vocabulary knowledge by consistently using the words in their academic English speaking and writing tasks [6].

In this paper, we proposed using the source and target text from NMT systems to generate personalized word lists to scaffold learning vocabulary. The experiment results demonstrated that our proposed idea provides students with a list of area-specific words that are mostly not found in other widely used academic word lists. This implies that the text in NMT systems can be analyzed to attain valuable insight into increasing specialized English vocabulary knowledge.

For future studies, we plan to generate and evaluate lists with a larger amount of text and for other subject areas. We also plan to generate more meaningful word lists by filtering out certain numbers and proper nouns. Furthermore, we plan to better understand our NMT system by comparing the differences between a word list generated from the source text and a list generated from a back-translation.