Emergence of Homonymy in an Evolutionary Language Game

Abstract

The evolutionary language game is an agent-based model in which agents communicate to develop a shared language over generations, each maintaining a phonetic vector representation of a set of disjoint concepts. While previous works have used perception rules and continuous vector representations, we compare different perception rules under varying phonetic vector dimensionalities to study the emergence of homonymy. Homonymy is multiple concepts using the same or similar phonetic representations, and we define it as cosine similarity ≥ 0.8. We simulate 100 homogeneous agents with 100 concepts over 20,000 generations and 16 seeds, sweeping phonetic dimensions 2–16. Agents begin with random vectors and update them under either argmax or threshold perception rules, with generation noise, transmission noise, and fitness-weighted turnover. We find that homonyms consistently emerge in lower phonetic dimensions with games using the threshold perception rule. By contrast, homonymy under argmax stays within a 1% relative difference of completely randomized vectors at all phonetic dimensions.  As the number of phonetic dimensions increases, alignment decreases in both perception rules. Thus, excess homonymy (homonymy greater than what is caused by randomized vectors) consistently emerges when agents are allowed to recognize a set of concepts rather than being required to identify a unique concept.