BACKGROUND: Simulation models (Plaut, 1996) and anomia treatment (Kiran and Thompson, 2003b) have found that training with atypical items resulted in greater generalisation than training with typical items. The Complexity Account of Treatment Efficacy (CATE) predicts that therapy will produce greater generalisation when more complex items are trained (Thompson, Shapiro, Kiran, and Sobeck, 2003). Furthermore, learning and generalisation may be maximised when semantic training is directed towards semantic deficits (Plaut, 1996). AIMS: The clinical application of Plaut's model and the CATE was investigated in typicality-based learning and generalisation, following anomia treatment in one participant with mainly phonological deficits and one participant with both phonological and semantic deficits. METHODS AND PROCEDURES: Participants were trained with a semantic treatment to name typical items in one semantic category and atypical items of a different semantic category. Typicality-based learning and generalisation were compared. OUTCOMES AND RESULTS: The participant with primarily phonological deficits learned typical items faster than atypical items, and showed no generalisation to untrained items. The participant with phonological and semantic deficits learned both typical and atypical items, and showed significant generalisation to untrained typical items and marginally significant generalisation to untrained atypical items. CONCLUSIONS: Our results suggest that typical items may be easier to learn than atypical items when the semantic system is intact; however, typical items may be more difficult to learn with impairments in semantic knowledge and/or semantic selection, as these items have highly similar semantic features. Additionally, correspondence between the deficit and treatment may enhance learning and generalisation, as semantic treatment resulted in greater learning and generalisation for the participant with semantic deficits. Finally, family resemblance within semantic categories may produce generalisation to typical or atypical category members, but training with atypical members may produce more global changes within the semantic network, leading to greater generalisation.