Training with Additional Semantic Constraints for Enhancing Neural Machine Translation

Replacing the traditional cross-entropy loss with BLEU as the optimization objective is a successful application of reinforcement learning (RL) in neural machine translation (NMT). However, a considerable weakness of the approach is that the monotonic optimization of BLEU’s training algorithm ignores the semantic fluency of the translation. One phenomenon is an incomprehensible translation accompanied by an ideal BLEU. In addition, sampling inefficiency as a common shortcoming of RL is more prominent in NMT. In this study, we address these issues in two ways. (1) We use the annealing schedule algorithm to add semantic evaluation for reinforcement training as part of the training objective. (2) We further attach a value iteration network to RL to transform the reward into a decision value, thereby making model training highly targeted and efficient. We use our approach on three representative language machine translation tasks, including low resource Mongolian-Chinese, agglutinative Japanese-English, and common task English-Chinese. Experiments show that our approach achieves significant improvements over the strong baselines, besides, it also saves nearly one-third of training time on different tasks.