We propose minimum risk training for end-to-end neural machine translation. Unlike conventional maximum likelihood estimation, minimum risk training is capable of optimizing model parameters directly with respect to arbitrary evaluation metrics, which are not necessarily differentiable. Experiments show that our approach achieves significant improvements over maximum likelihood estimation on a state-of-the-art neural machine translation system across various languages pairs. Transparent to architectures, our approach can be applied to more neural networks and potentially benefit more NLP tasks.
@article{arxiv.1512.02433,
title = {Minimum Risk Training for Neural Machine Translation},
author = {Shiqi Shen and Yong Cheng and Zhongjun He and Wei He and Hua Wu and Maosong Sun and Yang Liu},
journal= {arXiv preprint arXiv:1512.02433},
year = {2016}
}
Comments
Accepted for publication in Proceedings of ACL 2016