In defence of distributional analysis, pace Croft

In a number of publications (e.g. Croft 2001, 2004, 2006) Bill Croft has argued that distributional analysis as a methodology for setting up grammatical categories poses problems which can be avoided if constructions, not word classes, are grammatically primitive, and if categories are derived from constructions. He writes:

The Radical Construction Grammar analysis of parts of speech does not have Aristotelian grammatical categories of the sort envisioned by Aarts for particular language grammars. There are categories for each construction and each constructional role in a language. These construction-specific categories will have sharp boundaries to the extent that there are sharp acceptability judgements of what can and cannot occur in the relevant constructional role. In this sense, the categories are Aristotelian. But they do not lead to a small set of mutually exclusive word classes, which is what Aarts assumes we must posit. Instead, there are overlapping categories of formatives representing their diverse distributional behavior — which is what a speaker actually knows about her language. (Croft 2006: 10–11)

In this paper I will not be discussing the merits of Radical Construction Grammar, except to say that it is an interesting, challenging and exciting new approach to language. My aim here is more modest: I hope to show that distributional analysis is not as flawed as Croft suggests, and I will defend the distributional analyses discussed in Aarts (2004).