A residual-based approach to assessing conditional over- and under-representation of multiword expressions across phrase types

The present study introduces a residual-based approach for measuring the conditional over- and under-representation of two-word multiword expressions (MWEs) in language use. Using the British National Corpus, we model the expected frequency of an MWE from its constituent-word frequencies and phrase type. Each MWE’s residual, the difference between its observed and expected/predicted frequency, is interpreted as a descriptive indicator of over- or under-representation relative to this frequency-informed baseline. We also assess frequency projection, that is, how well constituent-word frequencies predict MWE frequency, and whether this projection differs across phrase types. Four phrase types were examined: adjective–noun, noun–noun, verb–noun, and verb–particle. Frequency projection was generally robust, though its strength varied across the four phrase types. The residuals were positively associated with PMI, t-score, and logDice, but the patterns differed across measures and phrase types. This suggests that deviation from expected frequency and co-occurrence strength capture related but distinct aspects of MWE use.