Bipartite spectral graph partitioning for clustering dialect varieties and detecting their linguistic features

Abstract

In this study we use bipartite spectral graph partitioning to simultaneously cluster varieties and identify their most distinctive linguistic features in Dutch dialect data. While clustering geographical varieties with respect to their features, e.g. pronunciation, is not new, the simultaneous identification of the features which give rise to the geographical clustering presents novel opportunities in dialectometry. Earlier methods aggregated sound differences and clustered on the basis of aggregate differences. The determination of the significant features which co-vary with cluster membership was carried out on a post hoc basis. Bipartite spectral graph clustering simultaneously seeks groups of individual features which are strongly associated, even while seeking groups of sites which share subsets of these same features. We show that the application of this method results in clear and sensible geographical groupings and discuss and analyze the importance of the concomitant features.