Connectionist models have generally not been able to perform natural language understanding or episodic memory retrieval beyond simple stereotypical situations that they have seen before. This is because they have had difficulties representing and applying general knowledge rules that specifically require variables, barring them from performing the high-level inferencing necessary for planning, reasoning, and natural language understanding.

This dissertation describes ROBIN, a structured connectionist model capable of high-level inferencing requiring variable bindings and rule application. Variable bindings are handled by signatures—activation patterns which uniquely identify the concept bound to a role. Signatures allow multiple role-bindings to be propagated across the network in parallel for rule application and dynamic inferencing. Signatures are integrated within a connectionist semantic network structure whose constraint-relaxation process allows it to perform disambiguation difficult for symbolic artificial intelligence models. These abilities make ROBIN an extremely promising approach to the problems of inferencing and natural language understanding.

We have also developed REMIND, a structured connectionist model based on ROBIN that explores the integration of language understanding and episodic memory retrieval with a single spreading-activation mechanism. The integration of episodic memory retrieval with ROBIN's high-level inferencing abilities gives REMIND a number of computational advantages over analogical and case-based retrieval models that perform retrieval only.

The influence of a high-level inferencing language understanding processes as proposed in REMIND has multiple implications on how human reminding works and how the two processes interact. In particular, it predicts that remote analogical reminding is possible through abstract thematic elements that the understander infers when processing language and that these episodic remindings in turn effect the comprehension process through priming. We describe collaborative cognitive psychology experiments that validate these predictions and imply that memory models need to be integrated with theories of language understanding, as predicted by some psychological researchers and explored in REMIND.