Open AccessError-driven versus batch models of the acquisition of phonotactics: David defeats Goliath
Author(s) -
Giorgio Magri
Publication year - 2014
Publication title -
proceedings of the annual meetings on phonology
Language(s) - English
Resource type - Journals
ISSN - 2377-3324
DOI - 10.3765/amp.v1i1.51
Subject(s) - phonotactics , restrictiveness , computer science , constraint (computer aided design) , artificial intelligence , ranking (information retrieval) , grammar , language acquisition , natural language processing , mathematics , linguistics , phonology , mathematics education , philosophy , geometry
Pure phonotactic learning is the problem of learning a restrictive OT grammar consistent with a set of licit surface forms. Hayes' (2004) LFCD and Prince and Tesar's (2004) BCD implement the batch approach to pure phonotactic learning. These algorithms are allowed to glimpse at the entire batch of data and can thus be endowed with aggressive biases towards restrictiveness. These authors dismiss the competing error-driven approach as computationally too weak to succeed at pure phonotactic learning. Indeed, an error-driven learner does not have access to the entire batch of data. Instead, it performs a sequence of re-rankings based on a single piece of data at the time. And its only provision towards restrictiveness consists of a restrictive initial ranking. In this paper, I compare the batch and the error-driven approach on two OT typologies yielding a total of roughly seventy languages. I show that the batch LFCD and BCD fail on a number of these languages. An error-driven learner which performs some constraint promotion besides demotion instead succeeds on each single language. Thus, David defeats Goliat.