Not all linguistic variation is equally predictable

Abstract We compare to what extent the choice of a variant can be predicted from language-internal factors for two linguistic variables: the English dative alternation and the omission of the infinitival marker att in a Swedish future-tense construction. Previous research has shown that for the dative alternation, near-ceiling performance can be achieved by fitting a regression model with manually selected predictors. Similar attempts for att -omission have been unsuccessful. To test whether the two variables differ in predictability or whether optimal methods have not been found for att -omission, we apply a large language model (LLM) to the same task. For the dative alternation, LLM and regression perform equally well. For att -omission, LLM outperforms regression, but still performs worse than for the dative alternation, thus suggesting that att -omission is inherently less predictable. We argue that LLMs can be useful for estimating how much the choice of a variant depends on language-internal factors.