next up previous
Next: Accounting for Graded Extraction Up: ExtractionGradedness, and Optimality Previous: Extraction from Picture NPs

Graded Optimality Theory


Optimality Theory (OT, cf. [Prince and Smolensky(1993), Grimshaw(1995)]) assumes that a grammar consists of ranked, violable constraints. In OT, the linguistic structure of an utterance is computed as the optimal structure of a set of possible candidates, where ``optimal'' is defined as satisfying most of the most highly ranked constraints. Like traditional linguistic theories, OT is based on a binary notion of grammaticality: the optimal candidate is grammatical; no assumptions are made about the relative grammaticality of suboptimal candidates.

To account for graded data, we propose to extend OT such that degrees of grammaticality are assigned to competing candidates according to the number and ranks of the constraints they violate. More precisely, we postulate that the optimality theoretic harmony of a candidate corresponds to its relative grammaticality in the candidate set. Under this assumption, evidence for constraint rankings can be obtained from judgments for suboptimal candidates. While a standard OT grammar makes predictions of the form: a structure tex2html_wrap_inline521  is grammatical, but tex2html_wrap_inline523  is ungrammatical, graded OT predicts that a structure  tex2html_wrap_inline521 more or less grammatical than  tex2html_wrap_inline523 , a claim that can be tested against graded experimental data.

Frank Keller
Thu Feb 13 14:35:39 GMT 1997