Bloem and La Heij (2003) reported that in a word-translation task context words induce semantic interference whereas context pictures induce semantic facilitation. This finding was accounted for by a model of lexical access in which: (a) semantic facilitation is localized at the conceptual level, (b) semantic interference is localized at the lexical level, and (c) only one concept is selected for lexicalization (the Conceptual Selection Model or CSM). In the present paper we report a second reversal: context words that induce semantic interference at SOA = +200ms, induce semantic facilitation at SOA = -400ms. The CSM simulated this finding when the additional assumption was made that lexical representations decay faster than conceptual representations. Experiments 2-5 test and confirm predictions of the CSM concerning the time course of semantic, phonological, and lexical context effects.