부정극어 허가 처리 평가: 영어 교과서를 학습한 L2 신경망 언어 모델을 활용하여

Abstract

The GRNN (Gulordava et al. 2018) neural language model (NLM) is viewed as a language learner in that it is trained with sentences and, like humans, ‘predicts’ the next-word given a sequence of words. Recent studies employing NLMs have reported their human-like performances in ‘understanding’ various linguistic phenomena. Building on previous studies, this paper aims to assess the level of linguistic knowledge that an NLM can acquire from a collection of English textbooks published in Korea for last two decades. We applied a psycholinguistic experimental method to compare the L2-GRNN to the L1-GRNN, focusing on the learning/processing of negative polarity item (NPI)-licensing in English. The L1-GRNN LM that was trained with the dataset from Wikipedia was reported to attain the linguistic knowledge that native speakers of English have. The L2-GRNN LM was trained with learning materials for Korean English learners. The result of analyzing the data extracted from the NLMs in NPI processing showed that the overall performance of the L2-GRNN NLM was far behind that of the L1-GRNN LM. In conclusion, this study demonstrates that the L2-GRNN has attained a far lower level of grammatical knowledge in NPI processing than its L1 counterpart. This result implies that the L2 dataset representing English textbooks published in Korea is not enough for the L2-GRNN NLM to acquire substantial grammatical knowledge that the L1-GRNN NLM as well as native speakers has.