Iterated language learning is shaped by a drive for optimizing lossy compression

Abstract

It has recently been theorized that languages evolve under pressure to attain near-optimal lossy compression of meanings into words. While this theory has been supported by broad crosslinguistic empirical evidence, it remains largely unknown what cognitive mechanisms may drive the cultural evolution of language toward near-optimal semantic systems. Here, we address this open question by studying language evolution in the lab via iterated learning. Across two qualitatively different domains (colors and Shepard circles), we find that semantic systems evolve toward the theoretical limit of efficient lossy compression, and over time, converge to highly efficient systems. This provides direct evidence that adult learners may operate under a bias to maintain efficiently compressed semantic representations. Moreover, it demonstrates how this bias can be amplified by cultural transmission, leading to the evolution of information-theoretically optimal semantic systems.