Continued development of Apertium-IceNLP: A rule-based Icelandic to English machine translation system
Ólafur Waage · 2010
We describe the continued development of the open source rule based Icelandic to English machine translation system (Brandt et al., 2010). It was created by combining the Apertium machine translation framework and adding the IceNLP toolkit with the purpose to increase the translation accuracy. It had an WER of 50.6% and a PER of 40.8%, scoring worse than the pure Apertium is-en language pair it was based on. We removed IceMorphy from the translation pipeline and replaced it with Apertium’s morphological analyzer (lt-proc). Then, we created a parser and mapper from Apertium’s stream format to IceNLP’s tag format for IceTagger to use. After this was completed, and after fixing the errors that cropped up along the way, the updated WER is 40.5% and PER is 21.3%. We conclude that increased accuracy might be achieved by looking deeper into Icelandic multiword expressions.