Constraint Conjunction versus Grounded Constraint Subhierarchies in Optimality Theory


A central assumption of classical Optimality Theory (Prince and Smolensky 1993) is that constraint domination is strict. That is, if constraint A outranks constraints B and C (A >> B, C), then a candidate output that violates A can never be favored over one that satisfies A but violates B and/or C, no matter how many violations of B and C occur. Informally, one might say that the number of times a candidate violates a constraint, or set of constraints, has no bearing on constraint ranking. One could imagine a different world. For instance, suppose A >> B, C unless both of the latter are violated. It is as though the violations of B and C together cause these constraints to be ‘promoted’ over A. In claiming that constraint violations do not add up in this way, Prince and Smolensky set Optimality Theory apart from the connectionist work that formed part of its inspiration.

But as is well known, Smolensky (1993, 1995, 1997) hypothesizes that just such an interaction of constraints exists, known as local constraint conjunction. Because it explicitly relinquishes the assumption of strict domination, even if only under some conditions, constraint conjunction represents an important departure from the more restrictive original conception of constraint ranking.

In this paper, I examine several significant uses of constraint conjunction that have been proposed, and suggest that they all can and should be subsumed under a notion independently required, that of the universal constraint subhierarchy. If this is correct, then constraint conjunction is not required of the theory. In addition, following Prince and Smolensky (1993), I take universal subhierarchies to be derived from linguistically relevant scales. Assuming these are phonetically or psycholinguistically grounded, then we have a promising means by which to address the challenge of overgeneration. The intuition guiding all of this paper is that apparent ‘worst of the worst’ effects are really ‘too much of a single bad thing’.

Download paper