Typed Feature Structures for Expressing and Computationally Implementing Feature Cooccurence Restrictions

It has long been part of the folklore that typed feature structures were introduced into linguistics in order to replace the perceived-to-be-messy FCRs of GPSG. This correspondence between FCRs and typed feature structures, however, has never been made explicit. In fact, with the notion of appropriateness conditions used in Carpenter (1992), it is initially diicult to see how any dependency could be expressed between the values of two features. We show in this paper that there is a way to express FCRs by expressing them on the maximally speciic types. These FCRs are unfortunately ignored in the type inferencing procedure from Carpenter (1992), which is used in the ALE parser of Carpenter (1993). We show, however, that by using the idea of \type resolution", it is possible to maintain satissable feature structures in parsing. This idea is implemented in the \Troll Type Resolution System".