INTRODUCTION The heart of any natural language processing (NLP) system is its underlying knowledge representation (KR) model. The limits of its KR become the limits of the entire NLP system. This is well known and accepted by researchers in symbolic systems (cf. Brachman, 1979; Woods, 1975). However, researchers in the various connectionist approaches to NLP are just beginning to come to grips with representing knowledge in artificial neural networks. When compared to their symbolic counterparts, connectionist NLP systems have woefully inade quate KR models. Typical shortcomings include the inability to represent the relationships between concepts in the model (“non-combinational semantics”) and the lack of a means of timely change in the system’s knowledge in response to the natural language input.