Sperber and Wilson's (e.g., 1986) relevance theory assumes different processing models for similar utterances without motivating the discrepancy (Giora, 1998). On some occasions, it seems to assume a direct access model on which the contextually appropriate interpretation (e.g., the figurative interpretation of metaphor) is accessed directly without having to process a contextually inappropriate interpretation, (e.g., the literal meaning of metaphor). On other occasions, it seems to assume a special sequential model on which the contextually inappropriate meaning for structure is involved in deriving the intended meaning. The graded salience hypothesis (Giora, 1997) may help reconcile the inconsistency. According to the graded salience hypothesis, salient (i.e., coded) meanings of words or expressions (whose degree of salience is affected by, e.g., frequency, familiarity, conventionality) and salient (e.g., frequent) structures should always be accessed and always first, regardless of contextual bias for the speaker's intent. According to the graded salience hypothesis, direct process should apply when salient information is intended, i.e., when salient information is compatible with contextual information. Sequential process should be induced when less salient meanings are intended (e.g., the literal meaning of conventional idioms). On such occasions, salient meanings would not be bypassed; rather, they would be activated first, rejected as the intended meaning and reinterpreted in harmony with the principle of relevance. Given the graded salience hypothesis, processes consistent with the principle of relevance may apply at different temporal moments of understanding, depending on the salience status of the discourse components involved. (LLBA 2000, vol. 34, n. 3)