Features, unification and nonmonotonicity

Feature structures and their associated unification operation have proved a very valuable tool for computational linguists. Lexicon writers, however, have expressed a need for nonmonotonic properties in their lexicons. Since traditional unification is monotonic, various researchers have come up with nonmonotonic variations on it. These variations invariably destroy other nice properties of the unification formalism, particularly its associative and commutative properties. In this paper we argue that the previous approaches to adding nonmonotonicity to unification are going about it the wrong way. Rather than try to replace unification with a nonmonotonic variation, it makes more sense to make a distinction between strict and default information, and keep unification as a simple information-combining operation. This allows us to define an associative and commutative unification operation that deals with nonmonotonic information. This paper introduces nonmonotonic structures and (default) logical semantics for them, and shows how some problems dealt with in other nonmonotonic formalisms may be translated into this one.