Honorific agreement and plural copying revisited: Experimental and deep learning approach

Honorific agreement and plural copying revisited: Experimental and deep learning approach

The study investigated two types of subjecthood diagnostics in Korean using two different kinds of approaches. Based on the previous studies on subjecthood diagnostics in Korean, this paper examined the two subjecthood diagnostics: Honorific Agreement (HA) and Plural Copying (PC). For the experimental analysis, this study adopted the analysis results of Kim et al. (2017). For the deep learning analysis, this paper employed the KR-BERT for the deep learning model and three sources of data sets: the Korean version of the Corpus of Linguistic Acceptability (K-CoLA), the Sejong Morphologically-Analyzed Corpus, and the extended target sentences of Kim et al. (2017). Two separate experiments were conducted in the deep learning analysis. In the first experiment, the KR-BERT was trained only with the K-CoLA, and the target sentences were analyzed. In the second experiment, the KR-BERT was trained with the K-CoLA and the sentences from the Sejong corpus, and the acceptability scores of the target sentences were measured. The acceptability scores were measured with the numeric scores using the algorithms in Lee (2021). After the experiments with the deep learning models, the scores were normalized and were statistically analyzed with Generalized Linear Models (GLMs). Through the two experiments, the following fact was observed: both HA and PC did not show similar tendencies of experimental results with human participants in the first experiment, but they did in the second experiment. The analysis results demonstrated that both HA and PC could be used as subjecthood diagnostics but that they played significant roles only when native speakers were exposed to enough examples.