Word embedding-based Part of Speech tagging in Tamil texts

This paper proposes a word embedding-based Part of Speech (POS) tagger for Tamil language. The experiments are conducted with different word embeddings BoW, TF-IDF, Word2vec, fastText and GloVe that are created using UJ-Tamil corpus. Different combinations of eight features with three classifiers linear SVM, Extreme Gradient Boosting and k-Nearest Neighbor are used to build the POS tagger. The results are compared against Viterbi algorithm-based POS tagger. The results show that word embedding can be used for POS tagging with good performance. BoW, TF-IDF and fastText give an impressive performance compared with Word2vec and GloVe. The accuracy of 99% is obtained with word embedding of BoW and TF-IDF with unigrams as well as bigrams and with linear SVM classifier. POS tag of a given word can be identified with 99% of accuracy using word embeddings based POS tagger in Tamil.