Data Augmentation by Concatenation for Low-Resource Translation: A Mystery and a Solution

In this paper, we investigate the driving factors behind concatenation, a\nsimple but effective data augmentation method for low-resource neural machine\ntranslation. Our experiments suggest that discourse context is unlikely the\ncause for the improvement of about +1 BLEU across four language pairs. Instead,\nwe demonstrate that the improvement comes from three other factors unrelated to\ndiscourse: context diversity, length diversity, and (to a lesser extent)\nposition shifting.\n