Abstract
In semantic categorization, nonwords that are neighbors of exemplars (e.g., turple in an animal categorization task) cause interference, but neighbors of nonexemplars (e.g., tabric) do not. This can be explained in a cascaded activation model in which the decision process selectively monitors activation in a category-relevant semantic feature unit. However, it is shown that this is true only for some categories. With the broad category “Physical Object”, interference is produced by nonwords based on both exemplars (e.g., himmer) and nonexemplars (e.g., travity). However, no interference is produced when the category is changed to “Animal”. This shows that only some semantic feature units can be monitored. It is proposed that what is being monitored are not in fact semantic features per se, but rather links to semantic fields defined on the basis of patterns of lexical co-occurrence.