(1) In feature theories of phonology, a term characterizing various approaches which see it as desirable that information should be omitted from underlying phonological representations. The representations should be minimally specified, or underspecified. There is a departure from the concept of ‘full’ specification present in early generative phonology: the view that the output of the phonological component must contain fully specified binary feature matrices. Underspecification theory is concerned with the extent to which feature distinctions should appear in a phonological representation, not as a binary choice of [+feature] v. [-feature], but as a choice between [+feature] and no marking at all. It therefore looks in particular at which feature values are predictable and may thus be left unspecified in a representation without harming the surface form.

The approach is chiefly associated with lexical phonology, but there are several underspecification models, which vary over their conceptions of minimality. In restricted or contrastive underspecification, only redundant features are lexically unspecified (e.g. in English, voicing would be specified for obstruents, where it is contrastive, but not for sonorants, where it is redundant). The approach limits the degree of underspecification in lexical forms by omitting only those feature values which are predictable on the basis of universal co-occurrence conditions. No other features may be underspecified. This contrasts with radical underspecification, which allows only one value to be specified in any given context in a representation. Moreover, such specifications are needed only when a rule would otherwise assign the wrong value to a feature. This approach omits from underlying representations not only the feature values which are predictable from co-occurrence conditions but also those which are predictable from context-free markedness statement. Default rules assign unmarked values. Other positions in underspecification theory are also possible, e.g. that the unmarked value is never introduced, so that all features are effectively single-valued (privative).

(2) The term is also used in relation to other levels of language for any model which does not require the specification of all the factors potentially involved in an analysis. In semantics, for example, there are approaches to formalization which do not completely specify all features of logical structure (e.g. in representing scope ambiguities).