First place for Copenhagen NLP in SIGMORPHON shared task – University of Copenhagen

17 January 2019

First place for Copenhagen NLP in SIGMORPHON shared task

Natural Language Processing

Yova Kementchedjhieva, Johannes Bjerva and Isabelle Augenstein from the NLP Group at Department of Computer Science, University of Copenhagen (Copenhagen NLP) won the ‘inflection in context’ subtask of the 2018 CoNLL-SIGMORPHON shared task.

CoNLL-SIGMORPHON is a shared task on multilingual morphological inflection, i.e., automatic generation of inflected word forms given a base form and a set of target morphological properties. The 2018 edition of the shared task presented a new subtask, where the target morphological properties are not given, but need to be recovered from context.

The Copenhagen NLP group won this subtask with a multi-tasking system, which simultaneously predicts both target forms and their morphological properties, the latter objective being trained in a multi-lingual fashion. 

Schematic representation of our approach, which includes a character-based neural attention function. The encoder has access to both the past and future context of each character.

You can read more about the approach in the publication: Yova Kementchedjhieva, Johannes Bjerva, Isabelle Augenstein. Copenhagen at CoNLL-SIGMORPHON 2018: Multilingual Inflection in Context with Explicit Morphosyntactic Decoding. Proceedings of ConLL--SIGMORPHON 2018

Yova Kementchedjhieva’s leading role in this project aligns with the overall topic of her PhD thesis, focusing on multilingual morphological processing with distant supervision. She is planned to continue the work carried out for CoNLL-SIGMORPHON 2018 during an internship with Google Research in the summer of 2019. The multilingual aspect of the project fits into Johannes Bjerva’s research, which is focused on multilingual learning in low-resource settings. Isabelle Augenstein is a tenure-track assistant professor in natural language processing and machine learning, who advises postdoc Johannes Bjerva and co-supervises Yova Kementchedjhieva’s PhD thesis.