Emergent Translation in Multi-Agent Communication

  • 2018-04-11 03:22:49
  • Jason Lee, Kyunghyun Cho, Jason Weston, Douwe Kiela
  • 0

Abstract

While most machine translation systems to date are trained on large parallelcorpora, humans learn language in a different way: by being grounded in anenvironment and interacting with other humans. In this work, we propose acommunication game where two agents, native speakers of their own respectivelanguages, jointly learn to solve a visual referential task. We find that theability to understand and translate a foreign language emerges as a means toachieve shared goals. The emergent translation is interactive and multimodal,and crucially does not require parallel corpora, but only monolingual,independent text and corresponding images. Our proposed translation modelachieves this by grounding the source and target languages into a shared visualmodality, and outperforms several baselines on both word-level andsentence-level translation tasks. Furthermore, we show that agents in amultilingual community learn to translate better and faster than in a bilingualcommunication setting.

 

Quick Read (beta)

loading the full paper ...