Although we are highly efficient at remembering and recognizing faces, we find it remarkably difficult to describe faces to others. This stems in part from the holistic nature of face representation and the difficulty of expressing configural information in words. However, research on dialogic interaction has shown that when interlocutors encounter a domain for which they possess only vague descriptions, they rapidly conventionalize ad-hoc ontologies that enable efficient and systematic communication. To test the hypothesis that interlocutors will conventionalize ad-hoc ontologies for describing faces, we present data from a collaborative, computer-based task which is played by pairs of participants. Successful completion of the task requires participants to interactively describe and identify target faces from a set of distractors. The data provide evidence of dyads conventionalizing ontologies for referring to faces. Drawing on the observed interactions, we demonstrate how these conventions emerge as a consequence of participants resolving instances of miscommunication.