POS Tagging Experts via Topic Modeling

Part of speech taggers generally perform well on homogeneous data sets, but their performance often varies considerably across different genres. In this paper we investigate the adaptation of POS taggers to individual genres by creating POS tagging experts. We use topic modeling to determine genres automatically and then build a tagging expert for each genre. We use Latent Dirichlet Allocation to cluster sentences into related topics, based on which we create the training experts for the POS tagger. Likewise, we cluster the test sentences into the same topics and annotate each sentence with the corresponding POS tagging expert. We show that using topic model experts enhances the accuracy of POS tagging by around half a percent point on average over the random baseline, and the 2-topic hard clustering model and the 10-topic soft clustering model improve over the full training set.