Population-genetic Thinking about Language Change

It is well-known that language change is a random process: we cannot predict either the course or the timing of changes with certainty. However, classical historical-linguistic models treat such uncertainty as a global-level phenomenon, without trying to formally tie it to the uncertainty at the level of individual speakers and linguistic tokens. Yet in many other empirical areas, efforts to provide microfoundations for global-level processes have been underway: economic behavior and traffic patterns are among the examples. Particularly successful were such efforts in population genetics: the study of the diachrony of genetic variation in populations.

In this talk I put forward a model of language change inspired by population-genetic methods. In that model, tokens of linguistic features are reproduced through time (i.e. repeated by speakers after being heard) in a random manner, so that sometimes a token may never serve as a model for a later token, and other times a token would be copied in later speech several times. Relatively rarely, speakers may misanalyze a token, and then reproduce their wrong analysis -- so that a miscopying, or mutation occurs. Certain types of linguistic features are communicatively more convenient in a given situation, and their tokens are copied more frequently, in effect giving rise to the linguistic analogue of natural selection. Speaker grammars pose constraints on possible reanalyses, and different speakers exchange with each other the linguistic features they produce in some network structure.

In this model, we can see that some known and puzzling patterns of language change arise under a wide range of conditions. Namely, the model predicts the existence of Sapir's drift -- the phenomenon when genetically related languages undergo similar changes, but long after they have separated, -- and of "unidirectionality with exceptions" in grammaticalization. Far from being mysterious, in the new model such patterns of change are among normal behaviors. This is due to the inherently stochastic nature of the explicitly modeled local-level process of change. That local-level stochasticity turns out to be enough to predict complex qualitative diachronic behaviors that didn't get a satisfactory explanation in the earlier literature.