When relearning words, patients with semantic dementia (SD) exhibit a characteristic rigidity, including a failure to generalise names to untrained exemplars of trained concepts. This has been attributed to an over-reliance on the medial temporal region which captures information in sparse, non-overlapping and therefore rigid representations. The current study extends previous investigations of SD relearning by re-examining the additional contribution made by the degraded cortical semantic system. The standard relearning protocol was modified by careful selection of foils to show that people with semantic dementia were sometimes able to extend their learning appropriately but that this correct generalisation was minimal (i.e. the patients under-generalised their learning). The revised assessment procedure highlighted the fact that, after relearning, the participants also incorrectly over-generalised the learned label to closely related concepts. It is unlikely that these behaviours would occur if the participants had only formed sparse hippocampal representations. These novel data build on the notion that people with semantic dementia engage both the degraded cortical semantic (neocortex) and the episodic (medial temporal) systems to learn. Because of neocortical damage to the anterior temporal lobes, relearning is disordered with a characteristic pattern of under- and over-generalisation. © 2011 Elsevier Ltd.