The Journal of Studies in Language ISSN:1225-4770(Print) 2671-6151(Online) 언어연구

Many diagnostics have been proposed to examine subjecthood in Korean, some of which were investigated in terms of their validity as subject diagnostics through a series of experimental studies. This paper takes the experimental data from three experimental studies (Kim et al., 2015; Lee et al., 2015a; Kim et al., 2017) that were previously conducted, and examines the validity of six diagnostic tests using cross-validation and a machine learning method. The experimental data were evenly divided into ten parts (10-fold cross-validation): Nine of them were used for machine learning (training data set), and the remaining one was for the testing (test data set). The analysis was conducted ten times, and the average values were taken. A simple linear regression was conducted for machine learning. Through the analysis, the following results were observed: (i) The accuracy of the diagnostics for Single Subject Construction (SSC) was much higher than that of Multiple Subject Construction (MSC) (SSC: 81,78, MSC: 68.66), (ii) Adjunct Control was the most accurate for SSC, whereas Coordinated Deletion for MSC, (iii) HA and PL Showed low performance. The results also seem to imply that some parts cannot be explained by a simple linear model but can be analyzed with non-linear modeling of language.