Bridging semantics and pragmatics in information-theoretic emergent communication

Abstract

Languages evolve through repeated interactions in rich contexts, where various communicative and non-communicative goals co-exist. The conveyed meaning is often shaped by the local conversational context of utterances, as captured by the pragmatic behavior of interlocutors, and at the same time, words are associated with non-contextualized meanings, as captured by lexical semantics. While semantics and pragmatics are widely studied, their interface and co-evolution is largely under-explored and not well understood. In this work we begin to address this major gap in our understanding by asking: How can a shared lexicon emerge from local pragmatic interactions? To this end, we build on a framework for information-theoretic emergent communication in artificial agents (Tucker et al., 2022). This framework is particularly relevant to our question because it integrates utility maximization, which is a central component in well-established models of pragmatics (Goodman and Frank, 2016), with general communicative constraints that are believed to shape human semantic systems (Zaslavsky et al., 2018). We adjust this framework to explicitly model the interface between semantics and pragmatics, such that agents learn to communicate in a pragmatic setting, i.e., in the presence of a shared conversational context, and then we evaluate their emergent lexicon. We test our model in a rich visual domain of naturalistic images, and find that human-like properties of the lexicon can emerge when agents are guided by both context-specific utility and general communicative pressures.