@inproceedings{lamar-kaya-2023-measuring,
    title = "Measuring the Impact of Data Augmentation Methods for Extremely Low-Resource {NMT}",
    author = "Lamar, Annie  and
      Kaya, Zeyneb",
    editor = "Ojha, Atul Kr.  and
      Liu, Chao-hong  and
      Vylomova, Ekaterina  and
      Pirinen, Flammie  and
      Abbott, Jade  and
      Washington, Jonathan  and
      Oco, Nathaniel  and
      Malykh, Valentin  and
      Logacheva, Varvara  and
      Zhao, Xiaobing",
    booktitle = "Proceedings of the Sixth Workshop on Technologies for Machine Translation of Low-Resource Languages (LoResMT 2023)",
    month = may,
    year = "2023",
    address = "Dubrovnik, Croatia",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/fix-sig-urls/2023.loresmt-1.8/",
    doi = "10.18653/v1/2023.loresmt-1.8",
    pages = "101--109",
    abstract = "Data augmentation (DA) is a popular strategy to boost performance on neural machine translation tasks. The impact of data augmentation in low-resource environments, particularly for diverse and scarce languages, is understudied. In this paper, we introduce a simple yet novel metric to measure the impact of several different data augmentation strategies. This metric, which we call Data Augmentation Advantage (DAA), quantifies how many true data pairs a synthetic data pair is worth in a particular experimental context. We demonstrate the utility of this metric by training models for several linguistically-varied datasets using the data augmentation methods of back-translation, SwitchOut, and sentence concatenation. In lower-resource tasks, DAA is an especially valuable metric for comparing DA performance as it provides a more effective way to quantify gains when BLEU scores are especially small and results across diverse languages are more divergent and difficult to assess."
}