Generalisation towards Combinatorial Productivity in Language Acquisition by Simple Recurrent Networks

Abstract

Language exhibits combinatorial productivity as complex constructions are composed of simple elements in a linear or hierarchical fashion. Complexity arises as one cannot be exposed to all possible combinations during ontogeny and yet to master a language one need to be, and very often is, able to generalise to process and comprehend constructions that are of novel combinations. Accounting for such an ability is a current challenge being tackled in connectionist research. In this study, we will first demonstrate that connectionist networks do generalise towards combinatorial productivity followed by an investigation of how the networks could achieve that