Learning in Harmonic Serialism and the necessity of a richer base

Abstract

This paper reassesses the hypothesis that early phonotactic learning of constraint-based grammars relies on the Identity Map – i.e. it uses observed surface forms as the inputs which cause errors and drive learning via constraint reranking. We argue that this approach's success is closely tied to Optimality Theory's fully parallel grammatical evaluation. In the constraint-based derivational framework of Harmonic Serialism (HS; McCarthy 2000, 2007b), reliance on observed surface forms as inputs can block the discovery of ‘hidden rankings’ between markedness constraints, preventing the learner from discovering a restrictive grammar. This paper illustrates the problem, using a pattern of positional vowel restrictions in Punu (Kwenzi Mikala 1980), and considers the role of various learning assumptions. We conclude that hidden rankings are a fundamental obstacle to restrictive error-driven learning in any HS-like framework, and that learning in such frameworks inevitably requires consideration of some unattested surface forms as inputs, even at the earliest learning stages.