Comparison of different POS tagging techniques for some South Asian languages

Abstract

There are different approaches to the problem of assigning a part of speech (POS) tag to each word of a natural language sentence. We present a comparison of the different approaches of POS tagging for the Bangla language and two other South Asian languages, as well as the baseline performances of different POS tagging techniques for the English language. The most widely used methods for English are the statistical methods i.e. n-gram based tagging or Hidden Markov Model (HMM) based tagging, the rule based or transformation based methods i.e. Brill’s tagger. Subsequent researches add various modifications to these basic approaches to improve the performance of the taggers for English. Here, we present an elaborate review of previous work in the area with the focus on South Asian Languages such as Hindi and Bangla. We experiment with Brill’s transformation based tagger and the supervised HMM based tagger without modifications for added improvement in accuracy, on English using training corpora of different sizes from the Brown corpus. We also compare the performances of these taggers on three South Asian languages with the focus on Bangla using two different tagsets and corpora of different sizes, which reveals that Brill's transformation based tagger performs considerably well for South Asian languages. We also check the baseline performances of the taggers for English and try to conclude how these approaches might perform if we use a considerable amount of annotated training corpus.