Automatic Lemmatisation for Norwegian
Proceedings of the Workshop on Structured Linguistic Data and Evaluation (SLiDE)
Abstract
We report on a new lemmatisation system for Norwegian, which is a particularly challenging language with two written standards, Bokmål and Nynorsk, that both have a lot of optionality. Our system covers both varieties and consists of a neural model that classifies words into rewrite rule classes that produce their lemma, as well as a large-scale computational lexicon of Norwegian that gives all possible inflections of a large part of the Norwegian vocabulary. We test different ways of combining these components. When evaluated with pure string-matching against the lemmas in the gold data, all systems perform approximately at the same level (99.1-99.2% on Bokmål and 98.5-98.6% on Nynorsk), but detailed error analysis shows that the computational lexicon reduces the number of true errors by more than half (reaching 99.6% accuracy on Bokmål and 99.3% on Nynorsk), as opposed to "surface errors" like using a different, but equally acceptable spelling variant of the correct lemma.