We studied the relationship between naming and semantic memory in a group of 10 patients with dementia of Alzheimer's type. In an extension to a previous cross-sectional study (Hodges, J. R. et al., Brain and Language, 1996, 54, 302-325), this relationship was investigated at two longitudinal points within each patient's cognitive decline. Two types of naming performance were compared: items that each patient named correctly at the first stage but failed to name at the second stage, as contrasted with items named correctly at both stages (thereby providing a control for cognitive decline in general). Semantic knowledge of the concepts represented by the pictures in the naming test was investigated at each stage using definitions to the spoken object name, scored particularly for the number of sensory and associative/functional features provided by the patient. At stage 2, an analysis of the definitions for named-->unnamed items as contrasted with named-->named objects revealed a significant loss of both sensory and associative information. A comparison between natural kinds (animals and birds) and artefacts (household objects, vehicles, etc.), however, demonstrated a striking interaction between category and type of information contained in the definitions. Specifically, stage 2 definitions of artefacts in the named-->unnamed set showed a disproportionate loss of associative/functional information, while definitions of animal names that patients failed to produce in response to the pictures were notably lacking in sensory features. This pattern supports the notion that successful naming relies on a subset of critical semantic features which vary somewhat across different categories of semantic knowledge. We suggest that these findings are best encompassed by a conception of semantic organization, Weighted Overlappingly Organized Features (WOOF), in which (i) knowledge about all objects is represented by a central, distributed network of features activated by both words and pictures, but (ii) natural kinds and artefacts are differentially weighted in favour of those features that are involved in learning about and experiencing different kinds of objects.