Learning Language: An Experiment

We develop a method for random assignment of language to participants in a controlled laboratory experiment, and use this to test the hypothesis that languages are learned more quickly when they can be identified with fewer number of observations. While the theory based on this hypothesis has generated substantial attention since being advanced by Blume (2005), evidence on its empirical validity has been elusive. Here we develop a novel extension of coordination games within which languages emerge endogenously. We show, first, that one can control features of an emergent language by varying the game’s incentives. This enables us to compare speed of learning across participants randomly assigned to different languages. Our data provide cogent evidence supporting the above hypothesis and Blume’s (2005) theory: Languages with compositional structures can be identified with fewer observations and are learned more quickly, and in this sense are efficient. Despite this, we find inefficient languages to sometimes emerge when they can be expressed using simple rules.