Why do we talk differently with different people? In this talk, Hawkins will argue for a dynamic account of lexical semantics that bridges psycholinguistics and sociolinguistics. Specifically, Hawkins will present a model of ad hoc lexical conventionalization via hierarchical Bayesian inference — using conversational feedback from partners to update one's beliefs not only about what is meaningful to them but also how meaning is indexed to broader social knowledge. Hawkins tests predictions of the model in two natural language communication experiments where participants are grouped into social networks for a referential task. Finally, Hawkins will discuss ongoing work exploring broader implications across three areas: (1) codeswitching and the construction of social personae, (2) developmental trajectories of pragmatic competence, and (3) artificial agents that can flexibly construct meaning with human partners on the fly. Overall, this line of work integrates insights from cognitive science, sociology, and machine learning to shed light on the dynamic, context-dependent nature of linguistic meaning.