Dynamic Curriculum Learning for Low-Resource Neural Machine Translation

• DOI: 10.18653/v1/2020.coling-main.352
• Publication Date: 2021-01-08T13:58:31Z
• Authors: Chen Xu; Bojie Hu; Yufan Jiang; Kai Feng; Zeyang Wang; Shen Huang; Qi Ju; Tong Xiao; Jingbo Zhu
• Keywords: Training set

Large amounts of data has made neural machine translation (NMT) a big success in recent years. But it is still challenge if we train these models on small-scale corpora. In this case, the way using appears to be more important. Here, investigate effective use training for low-resource NMT. particular, propose dynamic curriculum learning (DCL) method reorder samples training. Unlike previous work, do not static scoring function reordering. Instead, order dynamically determined two ways - loss decline and model competence. This eases by highlighting easy that current enough competence learn. We test our DCL Transformer-based system. Experimental results show outperforms several strong baselines three benchmarks different sized WMT’16 En-De.