Recurrent Neural Network Based Parts of Speech Tagger for Kannada Words

Each word, given its context and sense in a given sentence, can be assigned a POS tag. In natural language, a word may belong to more than one lexical category, so morphology of the word and its relationship with its neighboring words play an important role while assigning its proper tag. As a part of this study, Recurrent Neural Networks based Parts of Speech Taggers namely Bi-Directional LSTM, LSTM, RNN, GRU for Kannada words is proposed. The Enabling Minority Language Engineering Corpus containing 14,965 sentences with 23,044 manually tagged words is used to train the model. The accuracy achieved after applying the four proposed models namely, Bi-Directional LSTM, LSTM, RNN, GRU are 99. 23%, 98. 92%, 98. 19%, 99. 24% respectively.