A logistic regression approach to accent class division in Japanese dialects: With special reference to the Keihan-type accent system in peripheral Kinki regions

The accent system of Modern Japanese is characterized by patterns of high and low tones which are conventionally specified for each word, and words are grouped into several classes according to their accent patterns. Put differently, words belonging to the same class exhibit the same accent pattern. Some dialects, however, permit more than one accent patterns in the same class, which will be referred to as “accent class division” (ACD). In an attempt to explain this phenomenon, this paper takes a logistic regression approach, with the hypothesis that the features of consonants and vowels and their combinations are the possible factors. The data are from the Keihan-type accent in peripheral Kinki regions, especially Tokushima and Mie prefectures, where ACD is widely observed. It will be shown that the composition of segments in a word affects the choice of accent patterns, resulting in ACD.1