Training state-of-the-art portuguese POS taggers without handcrafted features

Abstract

Part-of-speech (POS) tagging for morphologically rich languages normally requires the use of handcrafted features that encapsulate clues about the language's morphology. In this work, we tackle Portuguese POS tagging using a deep neural network that employs a convolutional layer to learn character-level representation of words. We apply the network to three different corpora: the original Mac-Morpho corpus; a revised version of the Mac-Morpho corpus; and the Tycho Brahe corpus. Using the proposed approach, while avoiding the use of any handcrafted feature, we produce state-of-the-art POS taggers for the three corpora: 97.47% accuracy on the Mac-Morpho corpus; 97.31% accuracy on the revised Mac-Morpho corpus; and 97.17% accuracy on the Tycho Brahe corpus. These results represent an error reduction of 12.2%, 23.6% and 15.8%, respectively, on the best previous known result for each corpus.