Combining neural and non-neural methods for low-resource morphological reinflection

Abstract

We describe our systems and results in the type-level low-resource setting of the CoNLL–SIGMORPHON 2018 Shared Task on Universal Morphological Reinflection. We test non-neural transduction models, as well as more recent neural methods. We also investigate the effect of leveraging unannotated corpora to improve the performance of selected methods. Our best system obtains the highest accuracy on 34 out of 103 languages.