´ëÇѾð¾îÇÐȸThe Linguistic Association of Korea

ÇÐȸÁö

  • Ȩ
  • ÇÐȸÁö
  • ³í¹®ÀÚ·á½Ç

³í¹®ÀÚ·á½Ç

Á¦¸ñ Lexical Effects in Island Constraints: A Deep Learning Approach
ÀúÀÚ Yong-hun Lee
±Ç/È£ Á¦30±Ç / 1È£
Ãâó 179-201
³í¹®°ÔÀçÀÏ 2022-03-31
ÃÊ·Ï Lee, Yong-hun. (2022). Lexical effects in island constraints: A deep learning approach. The Linguistic Association of Korea Journal, 30(1), 179-201. This paper examined the lexical effects (a kind of random effect) of each experimental item in English island constraints. For this purpose, this paper adopted (i) the experimental design and dataset in Lee and Park (2018) and (ii) the deep learning model (the BERTLARGE model) in Lee (2021). After the BERTLARGE model was pretrained with the CoLA dataset, the acceptability scores were calculated for all the sentences in the dataset. As in Lee (2021), the acceptability scores in the BERTLARGE model were measured with the numerical values (neither TRUE/FALSE nor Likert scale), which was similar to the magnitude estimation in experimental syntax. After all the acceptability scores were collected, they were normalized into the z-scores and statistically analyzed. In this paper, a mixed-effects model was used where both fixed and random effects could be analyzed, but this paper focused on the random effects which were related to the lexicalization of experimental items. Through the analysis, the following was observed: (i) deep learning models could provide some help to make the experimental designs of syntax more sophisticated and fine-grained, (ii) it was possible to examine and control the lexical effects of experimental items with a deep learning model and a mixed-effects model, and (iii) in the case of island sentences, lexical variability was more crucially affected by the factor Island than Location.
ÆÄÀÏ PDFº¸±â  ´Ù¿î·Îµå