Analysis and development of Urdu POS tagged corpus
Ahmed Muaz, Aasim Ali, Sarmad Hussain · 2009
In this paper, two corpora of Urdu (with 110K and 120K words) tagged with different POS tagsets are used to train TnT and Tree taggers. Error analysis of both taggers is done to identify frequent confusions in tagging. Based on the analysis of tagging, and syntactic structure of Urdu, a more refined tagset is derived. The existing tagged corpora are tagged with the new tagset to develop a single corpus of 230K words and the TnT tagger is retrained. The results show improvement in tagging accuracy for individual corpora to 94.2% and also for the merged corpus to 91%. Implications of these results are discussed.