This study tested the predictions of the revised hierarchical (RHM) and morphological decomposition (MDM) models with Arabic-English bilinguals. The RHM (Kroll & Stewart, 1994) predicts that the amount of activation of first language translation equivalents is negatively correlated with second language (L2) proficiency. The MDM (Frost, Forster, & Deutsch, 1997) claims that in nonconcatenative languages, including Arabic, activation spreads by morphological identity rather than orthographic similarity. To test these two models, native speakers of Arabic at two levels of English L2 proficiency completed a translation recognition task. In the critical conditions, the Arabic word was not the correct translation of the English word ("shoulder-katif") but was orthographically related ("shoulder-kahf" "cave"), morphologically related but semantically opaque ("shoulder-takaatuf" "unity"), or semantically related ("shoulder-raqaba" "neck"). Results show more morphological- than orthographic-form interference for all participants, in line with the MDM. Contrary to the RHM, however, both proficiency groups experienced interference in the semantic condition as well as in the form conditions. (Contains 6 notes, 4 tables, and 2 figures.)