Lost in Translation: Addressing the ecological validity of agent-based models of communication emergence with insights from an empirical study

The agent-based approach to modeling language emergence recently attracts a lot of interest from the researchers from the AI community, who aim at recreating human-like communication skills in artificial systems. Importantly, such modeling also provides a useful tool for research in language sciences to tackle theoretical questions about the origins of the unique communication abilities in our species and conditions that gave rise to their structure.Within this work I argue that what hinders the field in progress towards achieving more natural human-like communication is maintaining too narrow a view of communication, focused primarily on modeling language as a static code responsible for conveying meanings. My main goal is to show how models could benefit from complementing the adopted perspective on language with greater recognition of the dynamics of interaction as a source of meaning construction.From this motivation I designed and conducted a study intended as a replication of a selected agent-based simulation in a real-world experiment with human participants. The idea was to ‘translate’ the digital setup into a physical one as accurately as possible in order to: 1) disclose any significant unnatural constraints imposed on the simulated social situation and flawed assumptions underlying the model; 2) gather observations on participants’ interaction and their communication strategies emerged in the context of performing a similar task as artificial agents. The obtained conclusions can not only help to improve the models by providing insights on how to simulate more ecologically valid social situations, but also by shifting the emphasis to outline more empirically informed desiderata for modeling more human-like communication.