Small Data, Big Impact: Leveraging Minimal Data for Effective Machine Translation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2740–2756 July 9-14, 2023 ©2023 Association for Computational Linguistics Small Data, Big Impact: Leveraging Minimal Data for Effective Machine Translation Jean Maillard∗ Meta AI Cynthia Gao Meta AI Elahe Kalbassi Meta AI Kaushik Ram Sadagopan Meta AI Vedanuj Goswami† Meta AI Philipp Koehn Johns Hopkins University Angela Fan Meta AI Francisco Guzmán Meta AI Abstract For many languages, machine translation progress is hindered by the lack of reliable train- ing data. Models are trained on whatever pre- existing datasets may be available and then aug- mented with synthetic data, because it is often not economical to pay for the creation of large- scale datasets. But for the case of low-resource languages, would the creation of a few thou- sand professionally translated sentence pairs give any benefit? In this paper, we show that it does. We describe a broad data collection effort in- volving around 6k professionally translated sen- tence pairs for each of 39 low-resource lan- guages, which we make publicly available. We analyse the gains of models trained on this small but high-quality data, showing that it has significant impact even when larger but lower quality pre-existing corpora are used, or when data is augmented with millions of sentences through backtranslation. 1 Introduction State of the art machine translation models are able to cover hundreds of languages (Ma et al., 2021; Wang et al., 2022; Siddhant et al., 2022; NLLB Team et al., 2022) by relying on large amounts of annotated (Skadin, š et al., 2014; Lison and Tiede- mann, 2016; Agi´c and Vuli´c, 2019) and unanno- tated web crawled data (Schwenk et al., 2021; Hef- fernan et al., 2022). Translation for low-resource languages still faces significant challenges related to data availability, since many of these languages have neither large-scale parallel corpora nor a big presence on the web (Adelani et al., 2022b). Techniques such as self-supervised learning (Ma et al., 2021; Liu et al., 2021) and backtranslation (Sennrich et al., 2016; Edunov et al., 2018; Fan et al., 2020) can be effective tools to reduce the ∗jeanm@meta.com † vedanuj@meta.com reliance on annotation for translation models. In some cases, these techniques can be combined or even be applied iteratively (Hoang et al., 2018), leading to a feedback loop that can generate in- creasingly better translations. In order to be effec- tive, however, such methods still require a certain amount of seed parallel data, which can be used to kickstart the process. As a result, researchers and communities looking to train translation systems for low-resource lan- guages may find themselves wondering how much parallel data is required to achieve a given perfor- mance target level. In this paper, we describe a data collection effort for 39 low-resource languages, involving the cre- ation of over 6k seed sentence pairs per language by professional translators, which we make pub- licly available with an open license. We analyse the behaviour of bilingual translation systems trained on …

Content truncated to 500 words