Detecting change and emergence for multiword expressions

Abstract

This work looks at a temporal aspect of multiword expressions (MWEs), namely that the behaviour of a given n-gram and its status as a MWE change over time. We propose a model in which context words have particular probabilities given a usage choice for an n-gram, and those usage choices have time dependent probabilities, and we put forward an expectationmaximisation technique for estimating the parameters from data with no annotation of usage choice. For a range of MWE usages of recent coinage, we evaluate whether the technique is able to detect the emerging usage.