Nine experiments of timed odd-even judgments examined how parity and number magnitude are accessed from Arabic and verbal numerals. With Arabic numerals, Ss used the rightmost digit to access a store of semantic number knowledge. Verbal numerals went through an additional stage of transcoding to base 10. Magnitude information was automatically accessed from Arabic numerals. Large numbers preferentially elicited a rightward response, and small numbers a leftward response. The Spatial-Numerical Association of Response Codes (SNARC) effect depended only on relative number magnitude and was weaker or absent with letters or verbal numerals. Direction did not vary with handedness or hemispheric dominance but was linked to the direction of writing, as it faded or even reversed in right-to-left writing Iranian Ss. The results supported a modular architecture for number processing, with distinct but interconnected Arabic, verbal, and magnitude representations.How are numbers mentally represented and manipulated? Despite recent advances in the cognitive psychology and neuropsychology of numerical abilities, theories of the basic architecture of number representations have remained highly controversial. Both modular and interactive views have been proposed to account for the same set of data. On the modular side, McCloskey, Caramazza, and Basili (1985; see also Mc-Closkey, 1992;McCloskey & Caramazza, 1987;McCloskey, Sokol, & Goodman, 1986) have postulated a central amodal and abstract representation of numbers, which would constitute a bottleneck entry to calculation routines and to stored number knowledge and that would be interfaced by notationspecific comprehension and production modules. On the opposite, interactive side, Campbell and Clark (1988; see also Clark & Campbell, 1991) have denied the existence of a central abstract representation and suggested that "visuospatial, verbal, and other modality-specific number codes are associatively connected as an encoding complex" (p. 204) and activate each other during retrieval and calculation. Somewhat intermediate models have been proposed by Dehaene (1992) and by Noel and Seron (in press). Dehaene's (1992) triple-code model postulates three cardinal representations of number-verbal, Arabic, and magnitude-each of which supports specific procedures such as number comparison or mental multiplication. Noel and Seron (in press), on the other hand, have hypothesized a unique preferred entry code, which would be used to access number knowledge and calculation routines but which may be either verbal or Arabic depending on the subject's idiosyncrasies.
