Predicting innovative alternations in Korean verb paradigms

In Korean verbal inflection, all forms in the paradigm (A-suffix, C-suffix, or ɨ-suffix forms) suffer from neutralization of some lexical contrasts, and there is no single form of the paradigm from which one can correctly predict all the other forms in the paradigm. Nevertheless, a survey of child errors and historical change show that the attested reanalyses are overwhelmingly based on ambiguities in A-forms, rather than in other affixal contexts (Kang 2006). In this study we conducted a computational modeling of learning of inflected forms of 952 Korean verbs using the Minimal Generalization Learner algorithm (MGL; Albright and Hayes 2002, 2003) to account for this striking asymmetry. The simulation result shows that A-suffix form correctly predicts the other forms in the paradigm at a higher rate than C-suffix form or ɨsuffix form does, indicating that the A-suffix form is indeed the most informative form of the paradigm. This is in line with the previous studies showing that learners designate the most informative form as a privileged base form (Albright 2002, 2008). We compare the model’s errors with attested child errors and historical changes.