Beyond lumping and splitting

Unlike synchronic linguists, who can ask their consultants for additional data and examples, historical linguists are, for the most part, stuck with the data we have. Only rarely are we able to add to the corpus of available texts, or find additional examples of a rare combination of phonemes. The problems of historical linguistics are thus inherently probabilistic: we must constantly decide whether perceived patterns in our data are significant or simply due to chance. Actually, this is true at any time depth; but the importance of probabilistic issues is especially clear when considering possible remote linguistic relationships. Unfortunately, the traditions of historical linguistics equip us poorly to handle these probabilistic issues. Our discipline had already achieved impressive scientific results in the nineteenth century, before modern significance-testing procedures were developed. Lacking reliable techniques for extracting maximal information from minimal data, and having plenty of other things to do, historical linguists have tended to focus on problems where meaningful results seemed assured, and have avoided investing time in riskier endeavours. Though the inherent interest and seductive appeal of the remote linguistic past are undeniable, pushing back the frontiers of time has generally not been a high priority for those actually working in historical linguistics. In this paper, we argue that the temporal reach of historical linguistics can probably be extended significantly in some cases, but only if the probabilistic issues are faced squarely, and new techniques of handling them are developed and widely understood. As an illustration, we propose a way of dealing with a Gedankenexperiment recently proposed by Hock & Joseph (1996, 491–3): if we had data only from Modern English and Modern Hindi, without evidence from other Indo-European languages, would we be able to detect the genetic relationship between them? Hock & Joseph express pessimism; but as we shall see, the problem yields to a fairly simple probabilistic technique. This example encourages us that our field will be able to penetrate deeper into the linguistic past.