Simultaneous Machine Translation using Deep Reinforcement Learning

We present a Deep Reinforcement Learning based approach for the task of real time machine translation. In the traditional machine translation setting, the translator system has to ‘wait’ till the end of the sentence before ‘committing’ any translation. Whereas real-time translators or ‘interpretors’ have to make a decision at every time step either to wait and gather more information about the context or translate and commit the current information. The goal of interpretors is to reduce the delay for translation without much loss in accuracy. We extend the work by (Grissom et al., 2014) and present a framework which uses the existing neural machine translation systems to function as a simultaneous machine translation system. We show that the policy learnt by our system outperforms the monotone and the batch translation policies while maintaining a delay-accuracy tradeoff.