Connectionism has been gaining ground as a psychological modelling technique annd shows great promise as a way to build fast, robust systems to perform intelligent tasks. Connectionism contrasts with the older tradition, that of explaining intelligent behaviour in terms of the manipulation of complex symbol structures. Without expecting either style of research to be exclusively correct, it is still important to seek out the relative strenghts and weaknesses of each. The partisans of the two camps have taken-up the challenge; the controversy has been fierce, as befits the prize-acceptance as the best technique for understanding people and building intelligent systems. Language, the quintessential human activity, is often used as a touchstone for the two approaches. The traditional approach views language as composed of symbols and language use as the manipulatoin of symbol structures. Connectionists have largely conceded to this, accepting the idea of structure as an essential component of language.This text calls into question this common assumption, that "structure is necessary for language modelling", by presenting a generator which produces appropriate natural language utterances without building structures along the way. It backs up this demonstration by an analysis of the generation task, which leads to the conclusion that massively parallel computation and numeric combination of evidence are, in fact intrinsically necessary for generation, and that, conversely, structure building is computationally awkward.
