The_Illiterati: Part-of-Speech Tagging for Magahi and Bhojpuri without even knowing the alphabet

In this paper, we describe the part-of-speechtagging experiments for Magahi and Bhojpuri that we conducted for our participation in the NSURL 2019 shared tasks 9 and 10 (Lowlevel NLP Tools for (Magahi|Bhojpuri) Language). We experiment with three different part-of-speech taggers and evaluate the impact of additional resources such as Brown clusters, word embeddings and transfer learning from additional tagged corpora in related languages. In a 10-fold cross-validation on the training data, our best-performing models achieve accuracies of 90.70% for Magahi and 94.08% for Bhojpuri. Accuracy increased to 94.79% for Magahi and dropped to 78.68% for Bhojpuri on the test data.