Representing grammar, meaning and knowledge

Abstract

Among the expertises relevant for successful natural language understanding are grammar, meaning and background knowledge, all of which must be represented in order to decode messages from text (or speech). The present paper is a sketch of one cooperation of grammar and meaning representations -- with some remarks about knowledge representation -- which allows that the representations involved be heterogeneous even while cooperating closely. The modules cooperate in what might be called a PLURALIST fashion, with few assumptions about the representations involved. In point of fact, the proposal is compatible with state-of-the-art representations from all three areas. The paper proceeeds from the nearly universal assumption that the grammar formalism is feature-based and insufficiently expressive for use in meaning representation. It then demonstrates how feature formalisms may be employed as a semantic metalanguage in order that semantic constraints may be expressed in a single formalism with grammatical constraints. This allows a tight coupling of syntax and semantics, the incorporation of nonsyntactic constraints (e.g., from knowledge representation) and the opportunity to underspecify meanings in novel ways -- including, e.g., ways which distinguish ambiguity and underspecification (vagueness). We retain scepticism vis-a-vis more ASSIMILATIONIST proposals for the interaction of these -- i.e., proposals which foresee common formalisms for grammar, meaning and knowledge representation. While such proposals rightfully claim to allow for closer integration, they fail to account for the motivations which distinguish formalisms - elaborate expressive strength in the case of semantic representations, monotonic (and preferably decidable) computation in the case of grammar formalisms, and the characterization of taxonomic reasoning in the case of knowledge representation.