Combining POS-taggers for improved accuracy on Swedish text

Several POS-taggers are trained and tested on Swedish text. Methods to improve the accuracy of the tagging are then examined. These methods include voting, letting taggers change their voting contribution depending on how confident they are and training a new second level classifier on the output of the taggers. All these methods are more accurate than the most accurate original tagger, with 15% less errors or better. Which types of errors these methods correct and which types remain are also examined. The number of errors in some common error categories actually increase, while many uncommon errors are corrected.