Current browse context:
cs.CL
Change to browse by:
References & Citations
Computer Science > Computation and Language
Title: On Compositional Generalization of Neural Machine Translation
(Submitted on 31 May 2021)
Abstract: Modern neural machine translation (NMT) models have achieved competitive performance in standard benchmarks such as WMT. However, there still exist significant issues such as robustness, domain generalization, etc. In this paper, we study NMT models from the perspective of compositional generalization by building a benchmark dataset, CoGnition, consisting of 216k clean and consistent sentence pairs. We quantitatively analyze effects of various factors using compound translation error rate, then demonstrate that the NMT model fails badly on compositional generalization, although it performs remarkably well under traditional metrics.
Link back to: arXiv, form interface, contact.