Simulated evolution of discourse with coupled recurrent networks

The origin and evolution of language have been the subjects of numerous debates and hypotheses. Nevertheless, they remain difficult to study in a scientific manner. In this paper, we focus on the string-context mutual segmentation hypothesis proposed by Merker and Okanoya, which is based on experimental findings related to animal songs. As a first step in formally exploring this hypothesis, we model the evolution of agent discourse using coupled recurrent networks (RNNs). This model is a simplified representation of this hypothesis; that is, agents are situated in a single context (e.g., behavioral, social: or environmental) and they mutually learn their utterance strings from the prediction dynamics of their RNNs. Our simulation demonstrates the emergence of shared utterance patterns, which are culturally transmitted from one generation to the next. Furthermore, the distribution of the shared patterns changes over the course of this evolution. These findings demonstrate an important aspect of language evolution: "language shaped by society.".