Uniformitarianism in diachronic typology

Question

Croft 2003 argues that "the typological universals discovered in contemporary languages should also apply to ancient and reconstructed languages" (the so called uniformitarian hypothesis, p. 233). How so? I don't quite follow his logic. How do new features emerge then? I'd appreciate some references.

Update: Let me borrow an example from Burlak and Starostin 2005. Let's say, we have three daughter languages, X, Y, and Z. Some feature F is present in all of those languages but it is realized by different means, F(x), F(y), and F(z). F(x) is the most typologically frequent form (based on "global" typology, not necessarily in this taxon), whereas F(y) and F(z) are less frequent. Now we need to reconstruct that feature for the parent language, W. What should we choose? Synchronic typologists would argue that it should be the most frequent form, F(x) in our case. Now there's the problem. If the parent language W indeed had the most frequent form F(x), then we'd have to argue that language states can change from a more typologically probable state to a less typologically probable one, a very undesirable result.

Answer 1

This is a very good question and I'm really happy I stumbled upon it. In general, however, I think there is here a very serious misunderstanding. The frequency of a typological feature is a statistical description of the occurrence of a certain phenomenon. Thus, we have the right to say that "since F(x) is more frequent than F(y) or F(z), it is more likely to have existed in W". But, and it is a very imporant but indeed, this does not in any way allow as to say that F(x) is really what W had. W is more likely to have had it, in the sense of probablistics, but we have no idea what it really had, in the actual linguistic sense.

Let me give you a small analogy: if I tell you I threw a coin five times and the last four throws it landed on tails, can you guess how it landed the first time? No. Maybe it landed on heads because that would be statistically more likely. But maybe it landed on tails because it's not perfectly balanced. All you can say is that the four heads in a row are a little odd and not likely to happen again soon if I use a proper coin.

One cannot draw conclusions about individual cases based on statistics. Statistics are constructed based on individual cases but they don't work backwards. They describe the result but say nothing about the cause. Statistics are a nice conclusion of a study but they're probably useless as a beginning. Most of the times, typology can't be used to solve a specific problem in historical linguistics.

To answer your question directly:

1. Croft 2003 might perhaps be right -- if and only if this conclusion follows directly from studies on reconstructed languages. But if I understand you correctly, he is trying to impose modern typology onto reconstruction. In that case his result may possibly be right but his methodology surely isn't. A bit like the ancient Greeks came up with atomism.

2. If frequency of F(x) &c. is all we have, then we simply have no way of telling what existed in W. Maybe some seemingly unrelated data can help; maybe there was strong external influence on Y and Z or maybe F(x) and F(y) have only been attested recently, while F(z) is very old? At any rate, more data is necessary because typology alone solves nothing here.

Answer 2

Imagine a simple system where there are two possible language types, A and B, and the total number of languages remains constant. Further assume that languages of type A change to type B with (stable) probability k(a), and languages of type B change to type A with probability k(b). When the system reaches a steady-state, the proportion of languages of type A will be

k(b) / ( k(a) + k(b) )

regardless of the relative frequences of types A and B at the initial state. A glance at a college chemistry book on first-order reactions might prove informative, but see Cysouw (2011) for a more thorough introduction to transition probabilities as typological variables.

Now imagine that k(a) is 4 times as large as k(b) (i.e. type A is diachronically unstable). This means that at steady-state, we expect for 20% of languages to be of type A, and 80% of languages to be of type B. This also means that on average, we should expect that for every language that changes from type B to type A, there are four languages that change from type A to type B. If we follow the uniformitarian principle, then there is no problem in proposing a transition from B-->A as a historical reconstruction, provided that B-->A transitions over languages are not proposed as being out of line with the proposed values of k(a) and k(b).