Minimizing UG: Constraints upon Constraints

This paper questions the validity of various assumptions central to treatments of featural processes such as vowel and consonant harmony, namely that features are universal, that constraints are universal, that feature scales are universal, and so on. The paper begins with a discussion of various phonetic, phonological, and typological properties of distinctive features. It goes on to discuss various properties of constraints, examining in particular the question of constraint plausibility. The class of anti-faithfulness constraints is examined in this regard, including a case study of Luo voicing polarity. It is customary in generative phonology to attribute a complex structure to Universal Grammar (UG), with numerous properties of natural language explained by the intrinsic structure of that grammatical component. For example, the sets of segments that exhibit class behavior are considered to be defined universally, with a single feature set appropriate for all languages (Chomsky & Halle 1968, etc.). Within Optimality Theory, such universality has been extended to the constraint set, with the postulation that phonological grammars draw on a fixed constraint set (Prince & Smolensky 1993, etc.). A major task for the linguist under such a conception is to uncover the properties of UG, for example, to discover the nature of the feature set and the composition of the constraint set. Given a ‘built-in’ set of constraints, for example, the problem for language acquisition consists of ranking the set of constraints but does not require the acquisition of constraints per se (Tesar 1995, etc.). Following work such as Hume & Tserdanelis (2002), Pulleyblank (2003), Mielke (2004, 2005), etc., I argue here that such a structure for UG is unlikely. I begin by considering certain properties of distinctive features, concluding that distinctive features are learned, not supplied by UG. It follows that if features are acquired, not given, then any constraints referring to features must also be learned, not precompiled. Certain implications for feature reference, constraint types and locality will be examined.