This repo is to replicate the model in the paper Deep Biaffine Attention for Neural Dependency Parsing.
Currently, on the CoNLL09 English Dataset, I fall behind 1-2 points on UAS/LAS.
CoNLL09-English | UAS | LAS |
---|---|---|
Valid | 93.52 | 90.16 |
Test | 94.22 | 91.68 |
Original | 95.21 | 93.20 |
Need more tuning for these models.
- This repo is based on my own framework bpase. You can git clone that repo and then
to install the library. There might be some dependency problems that I did not fix yet.
python setup.py install
- Quick Start
The model will be saved in saves.
python main.py
- Test
python test.py --trained_model saves/name_of_model
- Evaluation
Convert output to conll format with ConvertToConll.ipynb perl eval.pl -q -b -g results/conll.test -s results/conll.test.pred -o test.scores