December 8, 2019
Anti-efficient encoding in emergent communication
Neural Information Processing Systems (NeurIPS)
Despite renewed interest in emergent language simulations with neural networks, little is known about the basic properties of the induced code, and how they compare to human language. One fundamental characteristic of the latter, known as Zipf’s Law of Abbreviation (ZLA), is that more frequent words are efficiently associated to shorter strings. We study whether the same pattern emerges when two neural networks, a “speaker” and a “listener”, are trained to play a signaling game.
By: Rahma Chaabouni, Eugene Kharitonov, Emmanuel Dupoux, Marco Baroni
Facebook AI Research