pt-ud-diaparser

Biaffine dependency parser for Portuguese, trained with diaparser on UD Portuguese-Bosque and Portuguese-GSD (both CC BY-SA 4.0) with BERTimbau (neuralmind/bert-base-portuguese-cased, MIT) as encoder. It parses written words: each contraction (do = de + o, dá-lo) is one word with the head and relation of its part attached outside it, and a word cut at a hyphen or other punctuation (quarta-feira) links its pieces to the last one with the relation wp. Training script: train_dependency.py in this repo (--lang pt; also at scripts/train_dependency.py in Natural-Syntax-LSP), 30 epochs on a Colab T4.

Load with Parser.load("pt_ud.bertimbau") (trusted pickle; needs weights_only=False) or export to ONNX with export_model.py --mode dependency.

UD Portuguese-GSD test, all 1200 sentences, written words, punctuation excluded: UAS 92.6, LAS 89.0 (ONNX fp32); dynamic int8 UAS 92.1, LAS 87.8.