Assessing Nativelikeness of Korean College Students’ English Writing Using fastText

• Published in: 영어학 Vol. 22; pp. 19 - 39
• Main Author: Hyesun Cho
• Format: Journal Article
• Language: English
• Published: 한국영어학회 2022
• Subjects: 영어와문학; deep learning; fastText; plausibility; neural networks; grammaticality; nativelikeness; Korean EFL learners; English writing
• ISSN: 1598-1398; 2586-7474
• DOI: 10.15738/kjell.22..202201.19

Neural-network models have recently been used to assess nativelikeness of English sentences written by native or nonnative speakers. In this study, nativelikeness of Korean EFL college students’ English writing is assessed using fastText, a neural-network text classifier using subword information. The training data consisted of English sentences from the corpora of native speakers of English and Korean EFL college students. The test sentences consisted of English writing assignments written by Korean EFL college students. fastText performed well for the task of binary classification into native and nonnative sentences, with high accuracy in less than a minute. The sentences that are classified as native with a high probability tend to have fewer grammatical as well as plausibility errors than those classified as nonnative. For the test sentences, correcting grammatical errors (involving articles, number, subject-verb agreement, voice) had weaker effects on the classification of the sentences than correcting plausibility errors (word choices), which conforms to the previous literature. This suggests that fastText is more sensitive to plausibility errors than grammaticality errors which requires knowledge on hierarchical syntactic structures. KCI Citation Count: 1