Simultaneous Translation with Flexible Policy via Restricted Imitation Learning
This addresses the problem of flexible and efficient real-time translation for NLP applications, representing an incremental improvement over existing methods.
The paper tackles simultaneous translation by proposing a simpler single-model approach using a delay token and restricted dynamic oracle training, achieving better BLEU scores and lower latencies compared to fixed and RL-learned policies on Chinese<->English tasks.
Simultaneous translation is widely useful but remains one of the most difficult tasks in NLP. Previous work either uses fixed-latency policies, or train a complicated two-staged model using reinforcement learning. We propose a much simpler single model that adds a `delay' token to the target vocabulary, and design a restricted dynamic oracle to greatly simplify training. Experiments on Chinese<->English simultaneous translation show that our work leads to flexible policies that achieve better BLEU scores and lower latencies compared to both fixed and RL-learned policies.