UZH at CoNLL--SIGMORPHON 2018 Shared Task on Universal Morphological Reinflection

Abstract

This paper presents the submissions bythe University of Zurich to the CoNLL–SIGMORPHON 2018 Shared Task on Univer-sal Morphological Reinflection. Our system isbased on the prior work on neural transition-based transduction (Makarov and Clematide,2018b; Aharoni and Goldberg, 2017). Unlikethe prior work, we train the model in a fullyend-to-end fashion—without the need for anexternal character aligner—within the frame-work of imitation learning. In the type-levelmorphological inflection generation challenge(Task I), our five-strong ensemble outperformsall competitors in all three data-size settings.In the token-level inflection generation chal-lenge (Task II), our single model achieves thebest results on three out of four sub-tasks thatwe have participated in