Data Augmentation by Concatenation for Low-Resource Translation: A Mystery and a Solution
This addresses the problem of understanding data augmentation mechanisms for low-resource translation, though it is incremental as it clarifies an existing method.
The paper investigated why concatenation improves low-resource neural machine translation, finding that gains of about +1 BLEU across four language pairs are driven by context diversity, length diversity, and position shifting, not discourse context.
In this paper, we investigate the driving factors behind concatenation, a simple but effective data augmentation method for low-resource neural machine translation. Our experiments suggest that discourse context is unlikely the cause for the improvement of about +1 BLEU across four language pairs. Instead, we demonstrate that the improvement comes from three other factors unrelated to discourse: context diversity, length diversity, and (to a lesser extent) position shifting.