The LMU system for the conll-sigmorphon 2017 shared task on universal morphological reinflection

Abstract

We present the LMU system for the CoNLL-SIGMORPHON 2017 shared task on universal morphological reinflection, which consists of several subtasks, all concerned with producing an inflected form of a paradigm in different settings. Our solution is based on a neural sequence-to-sequence model, extended by preprocessing and data augmentation methods. Additionally, we develop a new algorithm for selecting the most suitable source form in the case of multi-source input, outperforming the baseline by 5.7% on average over all languages and settings. Finally, we propose a fine-tuning approach for the multi-source setting, and combine this with the source form detection, increasing accuracy by a further 4.6% on average.