Distributional approaches following the Firthian principle have revolutionized linguistics. While Firthian approaches in collocation research detect syntagmatic relations and are a key research area in corpus linguistics, Firthian distributional semantics and their neural counterpart of word embeddings detect paradigmatic relations and have fundamentally impacted computational linguistics. We combine these two closely related approaches: our hypothesis, following Ricoeur’s view of a metaphor as a clash of two normally distinct semantic fields, is that idioms are collocations in which the lexical participants typically have low semantic similarity in the word embedding space, i.e. low values for the cosine metric. We test if the cosine metric, replaceability with synonyms, and linear combinations with collocation measures improve idiom detection for three constructions: verb-PP, light verbs, and compound nouns. We report improved idiom detection by 10 to 80 per cent, and almost half of compound noun non-compositionality is predicted by cosine alone. We trace how compound nouns are changing in spoken and written English, mirroring digitalisation and the revolution of the internet.