Charles Explorer logo

Polishing the gold – how much revision do we need in treebanks?



We present the second version of PetroGold, a gold-standard treebank for the oil & gas domain in the Portuguese language. The corpus went through a series of revisions guided by three methods tested in the literature: inter-annotator disagreement, inconsistent n-grams and verification rules.

We perform an intrinsic evaluation and the model scores 90.92%, 89.09% and 84.07% in the UAS (unlabeled attachment score), LAS (labeled attachment score) and CLAS (content-word labeled attachment score) metrics respectively, CLAS being 1.11% higher than in the first version. We perform an experiment where we verify a negative impact in the intrinsic evaluation when simplifying the annotation related to prepositional verbal arguments and we conclude by discussing the results and future work.