It is well established that in masked priming, a target word (e.g., JUDGE) is primed more effectively by a transposed letter (TL) prime (e.g., jugde) than by an orthographic control prime (e.g., junpe). This is inconsistent with the slot coding schemes used in many models of visual word recognition. Several alternative coding schemes have been proposed in which special bigram detectors for frequently occurring nonadjacent letter combinations are developed as a product of perceptual learning. In order to examine this perceptual learning hypothesis, we asked whether bigram detectors are defined in terms of visuospatial coordinates. Japanese-English bilinguals who were equally familiar with horizontal and vertical text in Japanese demonstrated strong TL priming in both orientations when reading Japanese words, but, when reading English words, the evidence for vertical TL priming was not as strong. However, native English speakers showed a clear TL priming effect with vertically presented English words despite minimal exposure to vertical text, which is not consistent with a perceptual learning account. It is proposed instead that the initial letter array is transformed into an abstract ordinal code (first to last) regardless of orientation and that the speed with which this transformation is carried out depends on the familiarity of the script.