The Low-Resource Double Bind: An Empirical Study of Pruning for Low-Resource Machine Translation

A "bigger is better" explosion in the number of parameters in deep neural\nnetworks has made it increasingly challenging to make state-of-the-art networks\naccessible in compute-restricted environments. Compression techniques have\ntaken on renewed importance as a way to bridge the gap. However, evaluation of\nthe trade-offs incurred by popular compression techniques has been centered on\nhigh-resource datasets. In this work, we instead consider the impact of\ncompression in a data-limited regime. We introduce the term low-resource double\nbind to refer to the co-occurrence of data limitations and compute resource\nconstraints. This is a common setting for NLP for low-resource languages, yet\nthe trade-offs in performance are poorly studied. Our work offers surprising\ninsights into the relationship between capacity and generalization in\ndata-limited regimes for the task of machine translation. Our experiments on\nmagnitude pruning for translations from English into Yoruba, Hausa, Igbo and\nGerman show that in low-resource regimes, sparsity preserves performance on\nfrequent sentences but has a disparate impact on infrequent ones. However, it\nimproves robustness to out-of-distribution shifts, especially for datasets that\nare very distinct from the training distribution. Our findings suggest that\nsparsity can play a beneficial role at curbing memorization of low frequency\nattributes, and therefore offers a promising solution to the low-resource\ndouble bind.\n