A recent work from School of Interactive Computing faculty Dhruv Batra and Stefan Lee and collaborators Satwik Kottur and José Moura at Carnegie Mellon University has been recognized as the Best Short Paper at the 2017 Empirical Methods in Natural Language Processing (EMNLP) conference.
The work, titled “Natural Language Does Not Emerge ‘Naturally’ in Multi-Agent Dialog“, systematically explores the conditions under which human-interpretable languages emerge between interacting AI agents. In contrast to a number of recent works which have simultaneously found the emergence of grounded human-interpretable language between cooperative AI agents, this work finds that while most agent-invented languages are effective, they are decidedly not interpretable or compositional. In a sequence of ‘negative’ results culminating in a ‘positive’ one, this work demonstrates how it is possible to coax these invented languages to become more human-like and compositional by increasing restrictions on how the agents may communicate.
mlatgt.blog 