Croft (2001) argues that distributional analysis of word classes is doomed to failure because there is no way to know when to stop splitting word classes into subclasses. This paper discusses mathematical clustering algorithms and shows that contrary to Croft's assumption there exist hard and fast criteria to know when to stop splitting. The method exposed is applied to a subset of English lexemes first proposed by Crystal (1967). Finally, the clustering properties of typologically diverse languages are discussed in the light of the clustering model and checked against current theories of parts-of-speech. The paper concludes by affirming that clusterings can be established for any language but cannot be equated with the classical notion of parts-of-speech.