From Sequence Frequencies to Conditions in Bantu Vowel Harmony: Building a Grammar from the Ground Up

Abstract

We explore here the hypothesis that phonological grammars are emergent, formed by general principles that may involve little to no role for language-specific principles. Our basic proposal is that grammars develop from the identification of patterns of similarity, the calculation of frequencies of patterns of co-occurrence, and the development of generalized symbolic systems based on frequency data. We investigate six Bantu languages, all of which exhibit a canonical asymmetric height harmony pattern. Based on sizeable online databases, we examine the frequency of all possible vowel sequences in the six languages, using the frequency data to develop a nascent grammar for height harmony in each language. Our proposal is for a type of unsupervised learning, and we discuss various ways of establishing that the learning algorithm has converged on the correct grammar.