A Neural Solution to the Symbol Grounding Problem

Abstract

The Symbol Grounding Problem has been dealt with through a neural network architecture based on two interconnected modules: one deputed to symbol categorization, and another deputed to associate thematic roles to phrase components. The operation of the latter was implemented through the FGREP method (Forming Global Representations through Extended backPropagation), already introduced by Miikkulainen and Dyer [6]. Differently from what proposed by these authors, we used, as subvectors of vectors representing single phrase components, suitable codings of the categories to which the components themselves were associated by the categorization module. Such a modification let us obtain an improvement of performance of our architecture, with respect to the original Miikkulainen-Dyer model, in a task consisting of finding out the correct thematic roles of single components of phrases given as inputs.