The representation of world knowledge in a machine translation system is discussed, especially in terms of the interlingua text that is adopted in KBMT-89. While knowledge about the world is at the core of a knowledge-based machine translation program, it cannot be extracted from a text without knowledge about how languages encode it syntactically. The paper therefore also presents the grammar formalism with which language-specific syntactic knowledge is presented. Details are presented of our version of unification-based grammar, its basis in linguistic theory and its role in a knowledge-based machine translation system.