Seq2seq for morphological reinflection: When deep learning fails

Abstract

Recent studies showed that the sequence-to-sequence (seq2seq) model is a promising approach for morphological reinflection. At the CoNLL-SIGMORPHON 2017 Shared Task for universal morphological reinflection, we basically followed the approach with some minor variations. The results were remarkable in a certain sense. In high-resource scenarios our system achieved 91.46% accuracy (only modestly behind the best system by 3.85%), and in medium-resource scenarios the performance was 65.06% (almost the same as baseline). In low-resource settings, however, the performance was only 1.58%, ranking the worst among submitted systems. In this paper, we present system description and error analysis for the results.