Learning Phonology: Genetic Algorithms and Yoruba Tongue—Root Harmony

Abstract This chapter presents an implementation of a genetic algorithm which acquires the constraint rankings of an Optimality-theoretic grammar. The chapter begins by sketching the structure of a grammar within Optimality Theory, goes on to outline the nature of an acquisition model employing genetic algorithms, and finally, applies genetic algorithms to Optimality Theory, using as a case study the acquisition of tongue-root harmony systems. The basic structure of a phonological grammar in Optimality Theory is established by the appropriate ranking of a universally determined set of violable constraints. This structure is well suited for an acquisition model based on genetic algorithms, a form of evolution that occurs on a computer. In a genetic algorithm, a problem of some sort is encoded in the form of strings of symbols. In the case of an Optimality-theoretic grammar, each such string of symbols directly corresponds to a ranked set of constraints. A population of these strings (ranked constraint sets) is randomly generated, each corresponding to a possible grammar. The goodness (or ‘fitness’) of each of the strings is then assessed for how well it describes the sort of data encountered. The fitness of most of the randomly generated strings will be very poor—especially if the search space is large—but some strings are better than others.