Simultaneous Translation With Flexible Policy Via Restricted Imitation Learning
Baigong Zheng, Renjie Zheng, Mingbo Ma, Liang Huang . Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics 2019 – 74 citations
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.