Projecting Farsi POS Data To Tag Pashto

We present our findings on projecting part of speech (POS) information from a well resourced language, Farsi, to help tag a lower resourced language, Pashto, following Feldman and Hana (2010). We make a series of modifications to both tag transition and lexical emission parameter files generated from a hidden Markov model tagger, TnT, trained on the source language (Farsi). Changes to the emission parameters are immediately effective, whereas changes made to the transition information are most effective when we introduce a custom tagset. We reach our best results of 70.84% when we employ all emission and transition modifications to the Farsi corpus with the custom tagset.