Connectionism and natural language

Neural nets, have been successfully applied to various aspects of grammatical inference. The authors consider the relationship between neural nets and grammatical inference whilst taking account of some of the constraints involved in modelling natural language acquisition. These constraints are engendered by attempts to pay attention to the characteristics of the actual data upon which language acquisition is built. The authors explore these two issues: the sequential learning problem, and the notion of initial structure. They consider in more detail the reasons for giving neural nets a form of initial structure. A form that this initial structure could take is identified, and some recent research looking at the influence of initial structure on subsequent learning is described. The sequential learning problem is also described, together with some research which indicates what its probable solution is.<
>