Learning context-dependent categorization through a cooperative game

Categorization is an important process in cognition, and is still a topic of debate in linguistics, psychology, and philosophy. Historical descriptions of categorization have often relied on either Aristotelian, definition-based models, or prototype-based models in which categories are specified by referring to particular examples. We study a simple model of categorization, in which objects are represented as points in a multi-dimensional space and labeled by comparing them to neighboring points in a context-dependent manner. A group of agents learn categorization by playing a cooperative naming game, in which a speaker must communicate to a listener which of several objects is the target. By playing this game, the agents come to agree on a shared categorization scheme, which allows them to communicate effectively. We study how interactions between agents shape the emergent categorization scheme by quantifying the agreement of the agents with each other and with a standard color scheme. Our results highlight how the development of categorization depends on its usefulness not only in classification, but also in communication.