Despite the multiple semantic ambiguities present in every utterance during natural language use, people are remarkably efficient in establishing mutual understanding. This chapter illustrates how the study of human communication in novel settings provides a window into the mechanisms supporting the human competence to rapidly generate and understand novel shared symbols, capturing the joint construction of meaning across interacting agents. In this chapter, we discuss empirical findings and computational hypotheses generated in the context of an experimentally controlled non-verbal interactive task that throw light on these fundamental properties of human referential communication. The neural evidence reviewed here points to mechanisms shared across interlocutors of a communicative interaction. Those neural mechanisms implement predictions based on presumed knowledge and beliefs of the communicative partner. Computationally, the generation of novel meaningful symbolic representations might rely on cross-domain analogical mappings. Those mappings provide a mechanism for systematically augmenting individual pre-existing representations, adjusting them to the current conversational context.