A deep learning-based understanding of nativelikeness: A linguistic perspective

Abstract

Constructing deep learning models that identify nativelikeness in English sentences, this paper addresses two relevant research questions: is nativelikeness measurable, and is it determined by syntactic well-formedness and lexical associations? To address the first, our models are evaluated by judging every item in Test Suite I, which comprises learner and native sentences from four sources. The results show that the models predict nativelikeness reasonably well. Next, syntactic well-formedness is examined via Test Suite II, comprising correct–incorrect minimal pairs with two conditions. The results indicate that our models do not satisfactorily detect it. The learners’ results reveal their limited knowledge, suggesting that the models learn the inadequateness of lexical associations as a feature of non-nativelikeness because the learner training data comprises Korean English learner corpora. However, our models’ results also show poor performance. We conclude that deep learning is capable of measuring nativelikeness, and well-formedness and lexical associations are no more than necessary conditions for nativelikeness. This implies the need to consider other factors when defining and assessing nativelikeness. © 2021 KASELL All rights reserved.