This repository contains additional resources for the paper "Universal Dependencies for Tweets in Brazilian Portuguese: Tokenization and Part of Speech Tagging" to be published in ENIAC 2021.
We used a modified version of UDPipe 1, where we removed the parsing and tokenization components.
For UDPipe 2, we used the original repository located at here.
In Udify, we had to patch an update enabling the reading of treebanks without dependency relations. This is due to DANTE being under construction. Although, you should be able to use the original repo if your treebank has the dependency relations labels.
We evaluated model results with this script. You should be able to run it with the following command:
python3 udpipe2_eval.py gold_file.conllu pred_file.conllu --verbose
The confusion matrices for the final training is available in the assets
folder.