Katharina Kann (LMU Munich) “Low-resource Morphological Generation with Neural Sequence-to-Sequence Models”
3400 N Charles St, Baltimore, MD 21218
As languages other than English are moving more and more into the focus of NLP, accurate handling of morphology is getting constantly more important. This talk presents approaches to morphological generation, casting morphological inflection and reinflection as character-based sequence-to-sequence tasks. First, we will generally discuss how to successfully apply neural sequence-to-sequence networks to this type of tasks. Then, the focus of the talk will shift to the challenge that low-resource settings, which are unfortunately common for many morphologically rich languages, pose to neural models, which are known to require large amounts of training data. The approaches covered in this talk include multi-task learning, cross-lingual transfer learning and semi-supervised learning.
Katharina Kann is a third-year PhD student under the supervision of Hinrich Schutze at LMU Munich. During her PhD, she did research on different morphological tasks, and she won the SIGMORPHON 2016 shared task on morphological reinflection as well as more than half of the subtasks of the follow-up edition of the shared task in 2017. Since the beginning of October, she is interning at Google Zurich, where she is working with Katja Filippova on question answering
Previous to her PhD, she obtained her master’s degree in computer science from TU Munich in 2014 and her bachelor’s degree in mathematics from Johannes Gutenberg University of Mainz in 2011.