Abstract

Park, Myung-Kwan. 2024. Grammatical Transfer in L2 Neural Language Models. Sign Languages and Linguistics Convergence Research, 1(1), 58-86. This study examines the impacts of initial grammatical knowledge from the first language (English) on the acquisition of grammatical structures in the second language (Korean) by neural language models (NLMs). To this end, we employ the standard experimental procedure in the rapidly-evolving field of natural language processing (NLP) that encompasses (i) pre-training Transformer-based GPT-2 NLMs on either Korean or English datasets, (ii) further fine-tuning them using a targeted set of Korean data as either the first or the second language, and (iii) assessing their performances using the K-BLiMP test data while analyzing their grammatical generalization capabilities. The findings in this study indicate that negative transfer effects arise in L2 NLMs when comparing models that learn Korean as the second language with those that learn Korean as the first language. Additionally, during the progression of training, the NLMs learning the second language attained the acquisition of its grammatical features, but to different degrees depending on the construction types tested. This study enhances our understanding of potential grammatical difficulties that may occur due to the interference from the first language when NLMs learn Korean as the second language