Word-finding difficulty (anomia) is the most common linguistic deficit in dementia. It is often measured by picture naming tasks as naming a picture taps all the major processes in word production, i.e., activation of a concept, retrieval of lexical-semantic information on that concept, retrieval of the corresponding word form and articulation. Naming and naming errors have extensively been simulated by neural network models of lexicalization (see e.g. [1,2]). A common feature of these models is that they are static, i.e. non-learning. However, naming is a dynamic process that changes as a function of normal learning or re-learning after neural damage. These important patterns cannot be caught by the static models of lexicalization. Therefore we have developed a learning model of lexicalization based on multi-layer- perceptron (MLP) neural networks. We tested the model by fitting it to the naming data of 22 Finnish-speaking dementia patients and 19 neurologically intact control subjects. The tests showed an excellent fit between the model’s and the subjects naming response distributions. Thus our model seems be suitable to simulate naming disorders of dementia patients.