Disambiguation of Universal Dependencies Part-of-Speech Tags of Closed Class Words in Portuguese.

This paper explores methods to disambiguate Part-of-Speech (PoS) tags for closed class words in Brazilian Portuguese corpora annotated according to the Universal Dependencies annotation model. We evaluate disambiguation methods of different paradigms, namely a Markov-based method, a widely adopted parsing tool, and a BERT-based language modeling method. We compare their performances with two baselines, and observe a significant increase of more than 10% over the baselines for all proposed methods. We also show that while the BERT-based model outperforms the others reaching for the best case a 98% accuracy predicting the correct PoS tag, the use of the three methods as an Ensemble method offers more stable result according to the smaller variance for the numerical results we performed.