Modelling Differences in Multiword Expression Use Across Proficiency Levels in a Test of Spoken English

Abstract Despite the increased research on multiword expressions (MWEs) in the past two decades, there is a lack of theory that can explain the often-contradictory findings of those studies. Therefore, the aim of this paper was to propose two models, (a) one explaining differences in MWE density produced by speakers and writers at different proficiency levels and (b) another explaining differences in the difficulty of acquisition of different MWE groups. Taken together, they should theoretically predict differences in phrase density of MWEs produced by language users of different proficiency levels, with any L1-L2 combination, for various MWE groups.This paper validated the models for use for the test takers of the KIT Speaking Test (L1 Japanese-L2 English). The data used came from a 222,777-word corpus containing 574 examinees’ responses across four proficiency groups. Findings show that the proposed models predicted the phrase density patterns very well for 12 different MWE groups. These models not only contribute to theory, but also have the practical application of helping identify the MWE groups that raters should focus on to differentiate between test takers, including where lower rates of use should not be incorrectly attributed to a lower score.