In work on Feature Geometry, class nodes have the status of the sine qua non. An example of the well known phonological strategy whereby we seek representational solutions to problems, class nodes literally embody the underlying insight of feature classes. This paper attempts a fresh look at this area, and argues for a disembodied alternative, called Feature Class Theory, in which such classes are seen as nonrepresentational postulates of the theory. Much of the geometry is removed from a theory without class nodes, and features like Labial, Coronal, Dorsal and Pharyngeal, for example, are united in sharing the property of ‘placeness’ rather than in cohabiting under a node Place. That is, feature class theory postulates a set Place, of which these features are members. Constraints can mention a class like Place–hence the important idea of feature classes remains–but they thereby target the individual members of that set.
The occasional result of such individual feature targeting, and a central argument for the theory, is a certain type of partial class behavior, in which for example some, but not all, features of a class obey an imperative to spread, delink, etc.. Feature geometry theory, with its incarnation of feature classes as nodes, makes it difficult to see certain legitimate generalizations because they involve partial class behavior. This paper focuses on the proposed class Color, uniting the vowel place features [back] and [round] (Odden 1991, Selkirk 1991a). This grouping derives a great deal of support from patterns of vowel harmony, once we countenance the existence of partial class behavior like that illustrated here. Thus, I argue, feature class theory broadens the empirical and explanatory domain of feature classes to areas where feature geometry cannot go.
Crucial to an understanding of partial class behavior are the notions of constraint ranking and violability, and the work here draws heavily on Optimality Theory (henceforth OT, Prince and Smolensky 1993). A central claim here is that partial class behavior should be understood as the gradient violability of constraints targeting a feature class, often resulting from the constraint configuration C >> CONSTRAINT(CLASS), where C is any constraint, CONSTRAINT(CLASS) is a constraint targeting any class of features, and the two constraints conflict for some proper subset of the set Class.