Multi-agent models of language evolution usually involve agents giving names to internal independently constructed categories. We present an approach in which the creation of categories is part of the language formation process itself. When an agent does not have a word for a particular object it is allowed to use the existing name of another object, close to the original one as defined by an analogy function. In this way, the names in the shared lexicon that has evolved in a collective way, directly yield the different object classes. We present the results of several simulations using this model showing under what conditions the agents will develop meaningful classes. We also examine the effects of an influx and outflux of agents. Finally we discuss the prospects for models in which the classes would constitute relevant complex taxonomies.