Previous Seasons Meetings

What predicts productivity? Theory meets individuals 

Because they involve individual-level cognitive processes, psychological explanations of linguistic phenomena are in principle testable against individual behaviour. The present study draws on patterns of individual variation in corpus data to test explanations of productivity. Linguistic patterns are predicted to become more productive with higher type frequencies and lower token frequencies. This is because the formation of abstract mental representations is encouraged by varied types but counteracted by automation of high-frequency types. The predictions are tested for English -ly and -ness-derivation, as used by 698 individual journalists in the New York Times Annotated Corpus and 171 members of Parliament in the Hansard Corpus. Linear regression is used to model individual variation in productivity, in relation to type and token frequency, as well as several other predictor variables. While the expected effects are observed, there is also robust evidence of an interaction effect between type and token frequency, indicating that productivity is highest for patterns with many types and not-too-infrequent tokens. This fits best with a view of entrenchment as both a conservative and creative force in language. Further, some variation remains irreducibly individual and is not explained by currently known predictors of productivity.

The comparative method is one of the greatest methodological achievements in the history of linguistics. And yet, despite its relatively precise formulation, we do not have an automated implementation of it, and consequently we face a very long wait to know more about the inferable history of language families around the globe. One may well ask why. As it happens, in a mathematical PhD thesis from 2010, Alexandre Bouchard-Côté demonstrated why, by showing that even the inference of sound change was computationally infeasible. Bouchard-Côté pointed to two impediments: (1) a factorial explosion in the difficulty of the computational task, and (2) a paucity of evidence when the data consists of a short list of basic vocabulary. However, recent progress in computational statistics provides reason to believe that impediment (1) may be overcome for at least some models of linguistic change. Impediment (2) might be alleviated by allowing the algorithms to look at richer sources of data (as we humans do), such as inflectional paradigms. And so, in this talk I discuss the prospects for trying to automate a core aspect of the comparative method: the inference of sound change and analogy in paradigms, with an emphasis on analogy. I discuss what is already known about analogy; what it might entail to model that knowledge explicitly; the role to be played by mathematical models of language change; and what research questions the exercise might realistically help us to ask.