한국어 명사의 음소배열제약에 대한 기계학습

Machine learning of phonotactic constraints in Korean nouns

For the purpose of investigating not only gradient but also categorical phonotactic constraints in Korean, this study explores the machine learning of phonotactic constraints in Korean. In so doing, it focuses on phonotactic differences between native and Sino-Korean nouns. Employing 5,543 native Korean and 29,869 Sino-Korean words as the input training data, I ran a learning simulation, using a Maximum Entropy phonotactic model (Hayes and Wilson 2008). Based on the statistical distribution of the input data, markedness constraints were created with their own weights, the size of which reflects their gradient strength. The simulation results mostly confirm previous descriptions of phonotactics in Korean (for instance, no word-initial tense consonants in Sino-Korean). In addition, some previously unreported patterns were found