Statistical Remarks on Classification of Ainu Dialects : Lexicostatistics Revisited

There are two landmark lexicostatistical studies in Ainu dialectology focused on the classification of Ainu dialects, one by Hattori and Chiri (1960) and the other by Asai (1974b). However, two obstacles have prevented Ainu linguists from examining, comparing, and integrating Hattori and Chiri’s and Asai’s works. First, the two studies analyzed different lexicostatistical data and adopted different cognacy judgments. Second, the similarity among the Ainu dialects was measured using different methods in the two studies. This study focused on the second problem with the lexicostatistical data of Hattori and Chiri’s cognacy judgments. The main results of this study are summarized as follows: First, I reported miscalculations in Hattori and Chiri and proposed a corrected similarity matrix. Second, I examined two critical problems in lexicostatistics: the automatic quantification on “missing values” and the definition of numbers for language metrics. The discussions unfolded an alternative in lexicostatistics from Asai’s viewpoints. Third, I applied spectral clustering, a graph-theoretic approach adopted in previous studies (Ono 2020a, 2020b). Notably, the classification of Ainu dialects obtained statistically was generally consistent with those obtained by the same method to Asai’s data. Several differences between the two classifications were also reported in the disposition of the Horobetsu, Samani, Kushiro, Asahikawa, and Nayoro dialects. Fourth, I investigated the other strongest classifications in the Hokkaido Ainu dialects. The result led to an ABAtype distribution and its variant, which a previous study (Ono 2020a) also reported for Asai’s data using the same approach. The main results of this study suggest the need to reconsider not only Hattori and Chiri’s and Asai’s classification of Ainu dialects, but also previous lexicostatistical studies from methodological point of view.