Language versus dialect comparisonIn many ways, dialects are like related languages: in both cases, we are dealing with synchronic similarities, and an underlying history of divergence. One might therefore expect that any tools developed for classifying language families would generalize naturally to the dialect level. Yet in fact the preoccupations of classificatory historical linguisticsparticularly following the recent controversies over Greenberg (1987) and Dixon (1997), for instance -lie mainly with very large-scale comparison, and dialects tend to be ignored. Even if we did try to include dialects, the questions to be asked are not the same: we are not worrying about whether varieties are related, but how closely, and the usual models in historical linguistics simply do not work at the required level of resolution, or of detail.To illustrate this, consider what is perhaps the best known approximately quantificational method for calculating degrees of similarity between languages, namely lexicostatistics. This involves selecting some basic word meaning; identifying (or choosing -and this begs the significant question of just how we determine what the 'normal' translation is) its corresponding lexemes in a number of languages; and seeing if they are cognate. With the meaning 'castle', for instance, lexicostatistics would identify these corresponding forms in various Romance languages: Italian castello, Spanish Castillo, and French chäteau (fort); and all of these are derived from Latin castellum. It follows that a lexicostatistical quantification of the 'linguistic distance' between some Romance language varieties on the basis of this set of words would produce the outcome illustrated in Figure 1 below.Of course, Figure 1 is rather obvious, and might be thought a little unfair. However, it does make clear one most important point, which is that lexicostatistics is blind to what it scarcely takes a linguist to tell: namely that Italian castello [kas'tello] is far more like Spanish castillo [kas'tiXo] than either is to French chäteau [jato]. There is a great deal of comparative data in these words, awaiting quantification; but this cannot be achieved on