Collostructional analysis: a Bayesian approach

Abstract This article explores a Bayesian approach to collostructional analysis, a method that investigates the associations between lexical items and grammatical constructions. Traditional collostructional analysis has proved to be an effective tool for studying constructional semantics but has also faced several challenges, including a lack of cognitive plausibility and difficulties in selecting an association measure and in accounting for sampling variability. The proposed approach, based on a cognitively grounded Bayesian learning theory, implies measuring lexicogrammatical attraction by setting up a fully specified probabilistic model of construction–word relations. In a number of studies, both with fictional and real data, it is shown that the proposed method has several desirable properties. First, it frees the researcher from the necessity of choosing one among many available association measures or finding some ingenious way of combining them. Instead, association is interpreted simply and intuitively as the probability of observing a new instance of a lexical item in a certain construction. Second, as unobservable parameters are considered random variables in Bayesian modelling, my method efficiently solves two problems that have long plagued traditional collostructional analysis: sensitivity to sample size and difficulty in accounting for various levels of randomness in the data-generating process.