Two groups of Cantonese-speaking 5th graders were taught English new words and were later required to translate these English words into Cantonese (backward translation), and vice-versa (forward translation). One group learned the English words by associating them with their Cantonese equivalents (translation strategy) while the other group learned by associating them with verbal descriptions of their underlying concepts (concept strategy). It was found that the translation strategy produced shorter translation latencies and fewer errors than the concept strategy, and that backward translation was slower than forward translation when the translation learning strategy was used, thus questioning the hypothetical L2-to-L1 lexical link prescribed by the asymmetry model of bilingual memory (Kro11 & Stewart, 1994). The present findings are discussed in light of a concept mediation architecture for both lexicons (Chen & Leung, 1989; Chen, 1990; de Groot, Dannenburg, & van Hell, 1994; La Heij, Hooglander, Kerling, & van der Velden, in press).