언어 학습자로서의 신경망 언어모델: 영어 주어-동사 일치를 중심으로

Abstract

This paper assesses the language-processing ability of the neural language model (L2-GPT-2) trained on data sets of English textbooks published in Korea using the Generative Pre-trained Transformer (GPT)-2. Assuming that the language model (LM) is also an (artificial) language learner, we test it focusing on subject-verb agreement in English. It is a well-established fact that L1 speakers exhibit facilitatory interference effects in ungrammatical sentences with a plural subject and a singular form of verb. Unlike human native speakers, L2-GPT-2 as well as L1-GPT-2 display such effects in ungrammatical sentences either with a plural subject and a singular form of verb, or with a singular subject and a plural form of verb. Though there is a significant difference between human speakers and neural LMs in processing subject-verb agreement, the two LMs’ sensitivity to interference by a distractor NP points to the fact that they can attain a remarkably human-like linguistic generalization on subject-verb agreement.