Many models of word recognition predict that neighbours of target words will be activated during word processing. Cascaded models can make the additional prediction that semantic features of those neighbours get activated before the target has been uniquely identified. In two semantic decision tasks neighbours that were congruent (i.e., from the same category) or incongruent (i.e., from the opposite category) were presented in a long-term priming paradigm. Performance to targets was better if they were primed by congruent neighbours than if they were primed by incongruent neighbours. The same effect was found for rhyming and nonrhyming primes. The results support cascaded models that allow semantic information to become activated before lexical selection has finished. (Contains 2 tables and 4 footnotes.)