A Reinforcement Learning Approach to Interactive-Predictive Neural Machine Translation
This work addresses the challenge of minimizing human involvement in machine translation for users, though it appears incremental as it builds on existing interactive methods.
The paper tackles the problem of reducing human effort in interactive-predictive neural machine translation by using reinforcement learning from human judgments on partial translations, entropy-based uncertainty to trigger feedback, and online model updates, resulting in improved character F-score and BLEU with an average of 5 feedback requests per input.
We present an approach to interactive-predictive neural machine translation that attempts to reduce human effort from three directions: Firstly, instead of requiring humans to select, correct, or delete segments, we employ the idea of learning from human reinforcements in form of judgments on the quality of partial translations. Secondly, human effort is further reduced by using the entropy of word predictions as uncertainty criterion to trigger feedback requests. Lastly, online updates of the model parameters after every interaction allow the model to adapt quickly. We show in simulation experiments that reward signals on partial translations significantly improve character F-score and BLEU compared to feedback on full translations only, while human effort can be reduced to an average number of $5$ feedback requests for every input.