首页|Neural Translation System of Meta-Domain Transfer Based on Self-Ensemble and Self-Distillation
Neural Translation System of Meta-Domain Transfer Based on Self-Ensemble and Self-Distillation
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NSTL
Pleiades Publishing Inc
In order to address the shortcoming that feature representation limitation in machine translation (MT) system, this paper presents a transfer method for features in MT. Its main aim is to solve knowledge transfer of different training corpus in the decoding process. In this paper, the meta domain is modeled. A model agnostic self-ensemble and self-distillation training framework is proposed. The training is divided into model training and meta training to better adapt to the two types of features. In this paper, we have done extensive experiments on the classical neural machine translation system, and the model is compared with the classical methods. The experimental results show that the proposed model has improved in the transfer task of different domains and systems. In this paper, translation knowledge transfer is carried out on the Chinese-English translation dataset in the subdivided domain, which has a significant performance improvement in the news, education and law domain.
neural translation modeldeep transfer networkmeta domainself-ensembleself-distillation
NSTL
Pleiades Publishing Inc
